Repository navigation
0.7.0 integration: the answer-first stack reconciled with main - #710
Open
joyful-ii-V-I wants to merge 184 commits into
Open
joyful-ii-V-I wants to merge 184 commits into
joyful-ii-V-I wants to merge 184 commits into
Conversation
…ured
`--impact=intersection` answered `of="intersection" defs="1" reaches="2568"` and never
said WHERE that one definition is. Its file appeared only when a caller happened to live
in it; on textual's Region.intersection none does, so the blast radius of a change left
out the file the change is made in. A multi-group selector already lists its groups as
<def ix= t= [sc=] p=file:line/>; the one-group answer now prints that same row.
No new attribute and no new element: the compact legend's existing `def ix=` reading
applies, and --legend=full gets a one-sentence clause of its own (kSoleDefLegend), because
the multi-group clause describes of_def tokens no row of a one-group answer carries. Every
dialect of the verb: XML, columnar, --json def_table, MCP impact.
SCOPE: --impact only, by a stated reason. The radius of a change includes the file the
change is made in, so the definition is part of this verb's answer set. --callers,
--callees and --safe-delete keep the one-group answer without a table, as
attribmultidefcheck pins; impactsoledefcheck arm (6) holds that line.
MEASURED (lane t2-rankdefects, bar pre-registered; development data, the item-grain gold
is spent): on the 298 graded cells 211 -> 213 complete, 0 cells lost completeness, 282
cells byte-identical, cross-gold full pairs 253 -> 253 of 1,742. Over the 24 --impact
cells +1.84 % bytes against a 2 % bar; a row is 64 B, the rest is the legend clause, which
a --legend-have reader does not receive.
GATE, written first: test/impactsoledefcheck.sh, 19 arms, 8 of them red on the parent.
RE-PINNED, each by design:
* attribmultidefcheck d_one: the one-group --impact arm now pins exactly ONE <def> row
and still no of_def / prov on any row (plainSole).
* compactlegendcheck A-PIN impact.xml: previous pin equals the parent binary byte for
byte; the new one differs by the row and the clause, 4575 -> 4844 B.
* test/legend_over_target.tsv: the clause registered present-only on ripwire.impact/v1,
loop, mcp:impact (126 B) and full:impact (230 B).
* printf_parity.manifest: impact and impact_json, the only two labels that moved.
OWED AT INTEGRATION (not editable from this lane): a changelog.d fragment, and
`docs/counts_snapshot.py refresh` — this gate takes gates.scripts drift to 11 against a
bound of 10, and the refresh rewrites docs/TUNING.md, LIMITS.md and EVALS.md.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
… is 1 --impact lists the files that import a file defining SYM as <f> rows. Every one was printed <f via="import" p="…" lazy="0|1"/>. via="import" was a string literal of the one emitter that prints the row (serialize.h emitImportRowsXml), so it could not differ between two rows, and lazy="0" is the default arm of a two-valued flag. Counted before the change: 2,898 of 2,898 rows carried the literal and 2,844 of them lazy="0" (264 --impact answers: 24 graded cells and 240 drawn); the 100% rate is the source literal, not the sample. THE RULE IS CONTEXT-FREE. A row gains via="import" before p=, and lazy="0" when it has no lazy=. Nothing else is read, so the pre-change answer is rebuilt from the delivered answer alone and every grader returns the verdict it returned before. The --json dialect is untouched: a JSON reader has no tag to read the kind from, and its import_reach array is the gate's second reference. MEASURED on a draw registered before it was run (results/T3_LOSSLESS_1118, PREREG Amendment 2): 160 --impact answers over 20 shallow corpora, 1,597 importer rows. The <impact> element is 3.02% shorter and the whole answer 3.97% (45,866 of 1,156,210 B); 160 of 160 rebuild byte for byte, no answer is longer, 45 carry no importer row. On a frozen export of this repository 6.89% (n=12, 12 of 12 rebuild). The verb is 2.8% of the 298-cell graded population's bytes, so the population-level saving is about 0.14%. THE LEGEND. The compact reading of lazy=1 stays present-only, so it now prints only where a lazy row is; what a row WITHOUT lazy= claims is a new reading keyed on <f p=> under the impact schema, so an answer with no importer row pays nothing. bench/legendtier/census.tsv is amended by derivation and says so in place: the 22 of 39 answers that carried lazy on <f> are exactly the answers that carry an importer row, so they carry the new `p f` entry; `lazy f` is withdrawn rather than guessed (the census corpora are not in this repository; 13 of 264 answers carried a lazy="1" row on three other populations, under the one-third rule). Without that the impact pack fails legendtiercheck K3 at 3.19x (it priced a clause most answers no longer carry); with it, 2.84x. THE GATE (test/importrowshortcheck.sh, 38 arms plus an opt-in scale arm; 21 red on the unmodified binary) reconstructs the dropped bytes and compares the WHOLE <impact> element against two independent references: goldens captured from 994f002, and the --json import_reach array. Mutation controls (a deleted row, a changed path, a lazy flag added, one removed, a moved count outside the rows, the golden fed in as new output, an empty document) each go red for the named reason. Re-pinned, each by design: test/printf_parity.manifest (impact and help_all, exactly those two of 42), test/compactlegendfix/pre_l1_full/impact.xml (by hand, reversible by the same rule), and the row text eight gates grep for. compactlegendcheck D27/D30 read their lazy= row from test/impactimportfix, the one committed corpus with a lazy importer. Not done here: docs/COMMANDS.md and the dated captures still show the old row (docscommandscheck arm G is red at 994f002 before this change), and no changelog.d fragment is written. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…ss, 19.1 % of --affected bytes on the holdout
Owner directive 2026-11-18: the objective is the complete answer in the fewest bytes, and a LOSSLESS form is
approved even where it reverses E1 (2026-09-12, "every path verbatim", brace groups disqualified by name).
Paging the test list is not approved and is not done: every file is still named in the one call.
MEASURED FIRST (T4's 408 --affected answers, 27 repositories). The <g> lists are 48.3 % of the verb's bytes
and no path repeats inside an answer, so a table has nothing to index; what repeats is the directory and the
extension of sibling files. Share of all --affected bytes, census / holdout:
flat per-directory runs dir/{a,b} 12.5 % / 14.1 % (the earlier E1-blocked design)
per-answer directory table 13.5 % / 15.0 %
nested directories a/{b/{x,y},c} 16.4 % / 17.2 %
nested + extension a/{b/{x,y},c}.py 19.8 % / 20.7 % <- built
the same, any-character suffix 20.6 % / 22.2 % (declined: it splits file stems)
THE FORM
<g hops="2" n="4" pf="src/{a/{x,y}.test.ts,b.ts},README" run_each="node {p}"/>
[hops=2] (4 folded): src/{a/{x,y}.test.ts,b.ts}, README (run each: node {p}) (--situ)
THE READER'S RULE, which needs no ripwire: split on the commas outside braces; PRE{A,B}SUF stands for
PREASUF,PREBSUF in that order, braces nest. It is bash brace expansion over comma lists, and it yields exactly
the n= paths in the order p= listed them.
WHAT KEEPS IT UNAMBIGUOUS AND KEEPS SMALL ANSWERS STILL
- the row is RENAMED (pf=, "(n folded):"): p= / "(n):" are verbatim exactly as before, and a reader that
knows only p= fails loudly on a folded row, never wrongly;
- a list with a literal '{' or '}' in a path is never folded;
- only contiguous items are grouped, so evidence order is the expansion order by construction;
- a row folds only when that saves 8 B and a document only when its rows save 256 B in total; below the bar
the answer is byte-identical and its legend says nothing about the form;
- --affected, --test-gate and --situ opt in. JSON and every other verb on the row seam are byte-identical;
- tests=, shown_tests=, run_first=, n=, the commands and the exit codes are computed as before.
HOLDOUT (211 change sets, three verbs, 639 cells; base = this branch's parent):
--affected 719,387 -> 581,811 B -19.12 % folds 43 of 206, the other 163 byte-identical
--test-gate 997,534 -> 859,958 B -13.79 %
--situ 833,229 -> 688,837 B -17.33 %
0 exceptions to "unfold(new) == base byte for byte outside the legend"; 0 exit codes moved; two runs
byte-identical; one fired cell re-expanded by bash alone and compared with cmp.
NOT BUILT, BY MEASUREMENT: complement encoding ("every test file under R except these N"). A large answer
spreads its files over a median 18 evidence keys, so at most one row could be a complement: a 0.55 % ceiling
with the EXCEPT list priced, over a universe (what the crawl indexed and isTestPath accepted) a reader cannot
recompute from the answer or from the disk.
GATE, written and red before the source: test/affectedfoldcheck.sh. Three independent references — goldens
captured from the parent commit's binary (test/affectedfoldfix/golden), test/lib/pathfold.py (a reader and a
writer written from the rule, sharing no line with the C++), and bash's own brace expansion. Arms: unfold(new)
== golden for XML and --situ; the binary's fold == the reference writer's; what must not move (the
under-the-bar fixture, --json, --exercises, --pack-task, a literal-brace row beside folded ones); counts,
commands and exit codes; twelve mutation controls that go red; legend present-only; determinism; xmllint.
test/testrowpaths.py expands a folded row, so the gates that read through it are unchanged. Two private readers
learned the form: listingpagingcheck counts "(n folded):" as n files, and w3fixlegendcheck, which grepped full
paths in --situ's text, now reads them through the shared reader. The shard weight is affectedcheck's, borrowed,
not measured.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
`python3 docs/counts_snapshot.py refresh`, nothing hand-edited. affectedfoldcheck took the gate-script drift from the bound (10) to 11, so the per-PR check asks for the refresh in the same change. The other moved totals (caps, constants, tuning names) are drift that predates this lane. Kept as its own commit so an integrator who refreshes on the train can drop it. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…limited make -j test/lib/headbinlib.sh built the HEAD comparison binary with a bare `cmake --build -j`, which hands make an UNLIMITED -j: every translation unit of the ripwire target compiles at once, 1.3-3.4 GB per clang at -O3. On 2026-11-18 three such builds ran at once beside a runaway ripwire and the machine ran out of memory. headbin_jobs now picks the count: CMAKE_BUILD_PARALLEL_LEVEL, else RIPWIRE_HEADBIN_JOBS, else the logical CPU count under GITHUB_ACTIONS (the staging step runs alone on an idle runner, so CI keeps every core busy as before), else max(1, min(ncpu/2, GiB/8)). A set but non-integer override is ignored with one stderr line. bench/recalleval/run_r3diff.py's --jobs default was os.cpu_count(); it is now half the cores, at most 6. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…emand reference files An invoked skill loads its whole SKILL.md and keeps it in context for every later turn. The bodies were 261,872 B in total (median 9,478 B, max 41,668 B), so one average invocation cost 3-5x the entire per-session skill listing. Each SKILL.md is now the routing line, the decision, the one-call recipe, the stop rule and how to read the answer's honesty signals, plus an index of reference files in the same skill dir. Every other line of the old bodies moved verbatim into those files (flat *.md beside SKILL.md, the house pattern; installed by install(DIRECTORY skills/) and by install.sh's directory link/copy). SKILL.md total 261,872 B -> 52,931 B median / max 9,478 / 41,668 B -> 3,072 / 3,787 B frequency-weighted bytes per invocation (local verb telemetry) 26,254 B -> 3,553 B (-86%) Descriptions are unchanged: filler-only cuts saved 86 characters (1.7%) and cost 2.4 pp of dev-split bm25-desc routing (the swap trap again), so they were dropped before any held-out measurement. Measured under the pre-registered following check (33 tasks, blind sonnet + fable raters given only the skill text): correct first command 32/33 on both the old and the new bodies, per rater; the new bodies sent a rater to a reference file on 1 task of 33. Lexical routing: bm25-desc unchanged (split=test 64.6%), bm25-full judged 90 -> 94/160, for-routed judged 95 -> 100/160. Also fixes text that was already stale: the router's primer said quality-delta covers 10 kinds and that exit 2 means new debt; it covers 12 and exits 2 when code that already existed got worse. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…disclosed next= OWNER DECISION 2026-11-18. `--for` answers answer-first by default (item R, until now --legend=last) and stop after about 12 KB of rows. The coordinator's pricing on T6's re-run answers turned the single stop into PAGING: every page continues the same row stream at the same size, carries its own next=, and re-sends no legend. Pre-registered in results/AF1_ANSWERFIRST_1118/PREREG.md before any src/ change. WHY 5ee3ddb MADE IT OPT-IN, MEASURED. On the 154 design cells answer-first moves the cut/floor MARKERS into the prefix (first marker past 2 KB: 93/147 legend-first, 8/147 answer-first) and moves their READINGS out: a bare 12,000 B answer-first prefix shows a marker with no definition inside it in 137/154 cells (legend-first 0/154). A PAGE is not a bare prefix: page 1 keeps the WHOLE document legend after its rows and before </ctx>, and its own cut rides the ROOT (page= pages= withheld_bytes= next=), so the reader holds every reading of every marker it was served. That is the collapse condition 5ee3ddb wrote. THE MECHANISM (src/forpages.h). The answer is rendered whole, permuted answer-first, then cut at row boundaries (rowvalue's row_cuts granularity; never after a comment or a section open). Page 1 is the largest cut <= P from byte 0; page k the largest cut <= c(k-1) + P. A continuation page is the ref posture: a slim root, the re-opened section, its rows, and <about legend="ref" have="page1" .../>. est_tokens= prices each page as delivered in the emitter's two rates; scaling the whole answer's number UNDER-read by 571 tokens on formaxtokenscheck (B), because the bytes a page drops are not the answer's average mix. NOT A SECOND MECHANISM. --token-budget sheds (forbudgetfit.h) and is a different contract: the budgeted path keeps its legend-first shape unless --legend=last is SPELLED, and --page-bytes refuses beside a budget. The page vocabulary reuses next=, the ref posture's <about> and forUnbudgetedCall for the call every next= extends. OPT-OUTS. --legend=first is the legend-first order, unpaged: byte-identical to the base binary's default on 203/203 cells. --page-bytes=0 is the whole answer in one call: byte-identical to the base binary's --legend=last on 203/203. ho-llvm-01 is PENDING (the llvm memory defect lane M1 is fixing), not dropped. GATES. test/forpagescheck.sh (new, P1-P9): row identity of the joined pages against --page-bytes=0 at two sizes with a dropped-row negative control, no legend re-sent, next= chains, determinism, refusals. Re-pinned for the new default, each with a reason in place: foranswerfirstcheck (R9 -> R9 opt-out + R9d default), forrankordercheck (a reading must ride a legend comment of THIS answer, not only one ahead of <sigs>), budgetmonotonecheck (U is the whole legend-first answer a budget sheds), compactlegendcheck (UG) and legendcoveragecheck (an answer-first legend trails the rows), printf_parity (help text), flag counts 195 -> 196 (README x2, guide, deck x3), skill home for the flag. OWED, not done here: showcase capture + docs/COMMANDS.md regeneration (showcasecapturecheck (D) names --page-bytes uncovered), docs/gatecount_build.py (it writes docs/EVALS.md), EVALS amendment text, changelog.d entry, MCP paging. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tools/list is paid in every session by clients that load MCP schemas eagerly, before the first question. On the base (82ed4c8) it was 47,411 B by mcpmanifestcheck's metric (~11.9K tokens at 4 B/token, ~19.4K at 2.44). 17,365 B of that was description prose after the 33 pinned routing sentences. Trimmed only that prose: descriptions 22,498 -> 15,133 B (UTF-8), manifest -7,471 B (-15.8%). What went: clauses that repeated a property description, the answer's own legend (XML verbs), the readOnlyHint annotation ("read-only"), or described implementation: the short @file:LINE pointer (every field that takes a seed already says so), the JSON key lists, the exemplar ordering rule (still whole in the answer's legend; the "chosen by ROLE" phrase stays), "Same handler as the CLI", cost-model notes. What stayed: every floor/"not a proof" caveat, refusal and side-effect clause, band (1..12, 2..16, 1..1000), batch's served set and derived exclusion count, valueRefs, the tests_to_run row-shape clause for the legend-less JSON answers (304 -> 271 B, same facts), and the full @file:LINE sentence on find_symbol. Gate first: test/mcplisttrimcheck.sh pins every stanza minus its description (names, order, inputSchema incl. property descriptions, required, annotations) against test/mcplist_structure_golden.txt taken from the base binary, a 39,970 B ceiling, two-server determinism, and a mutation control. RED on the base binary (ceiling arm), GREEN here. mcpmanifestcheck's ceiling ratchets 47,420 -> 39,970 and its (1c) whereis twin moves 701 -> 545. Registered in regression.sh, shard weights and gateexitcheck FAILFAST. Owed (not touched by this lane): docs/counts_snapshot.py refresh (gates.scripts 852 -> 853; it rewrites README, docs/EVALS.md, the deck and docs/counts.tsv), and a changelog.d fragment. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…f (RIPWIRE_MEM_BUDGET_MB) On 2026-11-18 a run over llvm-project grew ~5 GB/s inside buildGraph's resolve loop and reached 149.7 GB. The memory guard (memguard.h, 65% of RAM) reads the footprint only at phase boundaries, in the crawl/parse workers and in the file-graph fold; nothing between "graph-tables" and "graph-end" reads it, and one call can grow without bound. src/memceiling.h reads the footprint at operator new (largealloc.cpp, macOS): every allocation of 256 KiB or more, and once per 64 MiB of smaller allocation per thread. Past the ceiling it refuses: one stderr line from a stack buffer (no allocation inside operator new), nothing on stdout, std::_Exit(5) -- 5 is the memory refusal everywhere. Disarmed it costs one relaxed load per operator new. --mcp/--lsp and builds that do not replace operator new do not arm it and say so on stderr; a malformed value refuses the run (exit 1), as RIPWIRE_MAX_MEMORY does. memstats::mark records the last phase boundary for the refusal line. Proof, run once alone under the watchdog: llvm-project with RIPWIRE_MEM_BUDGET_MB=4000 refused with exit 5 at a 3018 MiB footprint, before a 1760 MiB allocation, after phase "model" (watchdog peak RSS 2723 MB). test/memceilingcheck.sh: refusal shape and exit 5 at the floor, the unset control, byte-identical output armed far above the need, typo refusals, and an MCP server that stays up under the variable. infra/envnumber.h: the environment whole-number reader RIPWIRE_TAIL_SYM_WIDE_ROWS and the new knob share. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… or a signal-kill as its own state A runaway ripwire with no timeout ran until macOS killed it, and the bench scripts read the result as an empty answer or a generic error. The whole-repo spawns in bench_proof, scip_amb_precision, scip_match_diag, scip_pin_precision, svectorab, timing_determinism, nestcal/measure, slice/run_slicerecall and legendtier/census now carry a timeout, and a timeout or rc < 0 is reported by name instead of being counted. MCP sessions (mcp_session_timing, spec_trace) read each reply through one bounded reader that raises TIMEOUT or KILLED. A normal run's output is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ied to the default page size Built from results/AF2_FIRST12K_1118/READOUT.md §SPEC (frozen after meeting its pre-registered bar: within 12 KB, design 70->75 of 99, held-out 32->33 of 39, 0 flips). Order only: every byte of the answer is still emitted once. A A comment that is a direct child of a top-level section (<graph_verb>, <reexports>, <fold> notes) joins the document legend; a zero-row self-closing <lego/>/<compose/> stub between <sigs> and <tail> goes directly after </tail>. A populated lego/compose does not move. E1 Inside <graph_verb>, <met_bridge> is emitted ahead of the <path_gap> sides (an emitter swap). G GUARDED promotion: a RES section after </tail> (lit, dotted_strings, graph_verb, callers, selfhooks, manifest_rows, listed, ep, reexports, exemplar_family) moves directly ahead of <tail> only if the tail has more than 24 rows, ends past the stop, and tail row 24 still ends within the stop with it in place. The stop is kAnswerPageBytesDefault, the ONE default page-size constant that forpages.h's default IS, so the guard follows the owner if the default page becomes 6 KB. It is not per call: every page size windows the same document, so --page-bytes=0 stays the row-identity reference for all of them. The unguarded variant is an owner call and is not built. ONE DEVIATION FROM THE SPEC, and why. The spec also lifts <sigs>'s demote:/decl-twin:/one-side: comments. Those are POSITIONAL: each labels the <t> pointer rows that follow it, and fordemoteallcheck read rows 9-10 as decl-twin and 11-26 as demote from exactly that order. Lifted into the legend, every <t> reads as either kind -- fordemoteallcheck (d)/(j)/(x), fordecltwincheck, foronesidecheck and forsiblingdemotecheck all went red on it. A lost disclosure is not a byte saving, so <sigs> keeps them in place; the readout measures what that costs against AF2's figures. GATES. foranswerfirstcheck: R1/R2 now judge AF2's UNIT multiset (open/close tags + direct children) and the comment COUNT over every unit, R1n still fails on a dropped row; new arms (A1) no section but <sigs>/<bodies> keeps a direct-child comment, (A2) no zero-row stub ahead of <tail>, (G1) a promoted section leaves tail row 24 inside the stop with </tail> past it, (G2) nothing moves when </tail> is inside the stop, (G0) the literal fixture must FIRE G. Against the core binary A1/A2/G0 go red (controls). forpagescheck (P10): the flagless default is --page-bytes=12000's page 1. Re-pinned: gvassemblycheck (e1) and graphformendpointscheck (f1) compare the embedded verb answer without its lifted comments, (b1)/(g1) read the bridge/gap legend from the document legend and expect the bridge first; foronesidecheck (t) takes the whole legend-first answer as the U a fitting budget returns, as budgetmonotonecheck (a) does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…to the memory guard's limit Owner, 2026-11-18: "On by default". The ceiling is now the memory guard's hard limit (65% of min(RAM, cgroup memory.max), --max-memory, RIPWIRE_MAX_MEMORY: the flag wins over the variable as before), read at operator new: one knob, one exit code (5). RIPWIRE_MEM_BUDGET_MB=<MiB> sets this line alone; RIPWIRE_MEM_BUDGET_MB=0 turns it off. The refusal names the knob that set the limit. Portable hook: src/memceiling_new.cpp replaces operator new (malloc/free plus the hook) wherever largealloc.cpp is not linked — Linux, Windows, other POSIX; never beside a sanitizer or RIPWIRE_ALLOC_COUNT, which leave the line disarmed (RIPWIRE_MEM_CEILING_HOOK undefined). The readers are rw::os::mem_footprint: phys_footprint, /proc/self/statm, GetProcessMemoryInfo commit charge. RIPWIRE_STATS prints the line's reading count at graph-end. An operator-new allocation can no longer pass the limit between two phase readings, so the overshoot memguard disclosed is now refused for that memory; what the line cannot see (direct malloc, mmap, the small-allocation stride) is still disclosed. largeallocpeakcheck arm M, red on the parent commit too, passes now. memguardcheck B8s/B8bs and B12 assert memguard's own phase and parse lines, so they run with the line off; B8/B8b accept its refusal as a real line. test/memceilingcheck.sh: default-on, both knobs, the 0 override, byte-identical map/--json/--for with the line on, off and armed high, typo refusals, and an MCP server that stays up. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eclarations they measured
`--uses=patch_vary_headers` answered `defs="2"` and never said where either definition is; a
one-group --callers answer did the same, and a one-group --impact answer (T2's IMPACT-SOLE-DEF)
named the definition but not the header that declares it. Lane P1 found that 13 of the 16 gold
groups a routed --callers/--uses/--impact answer lost were the file that DEFINES or DECLARES the
symbol (results/P1_VERBROUTE_1118, SPEC §1).
THE RULE (pre-registered, results/V1_VERBDEF_1118/PREREG.md):
* --callers (XML, columnar, --json), --uses (XML, columnar) and --impact, and the MCP twins
find_referencing_symbols, uses and impact, print the <def> table on every answer whose
selector resolved a definition group, one group included. --callees, --safe-delete and
find_symbol keep a one-group answer table-less.
* Each listed group's row is followed by one row per declaration tied to its definition (the
existing identity rule: C family edit-check, TypeScript overload signatures), same ix=,
bodyless="1". A header prototype is now named beside the definition (tally: proto.c:3 then
proto.h:3). This applies to every <def> table, so one reading holds tool-wide.
* Group listing unchanged (16-group cap, defs_listed=N). --uses rows name no group, so above
16 groups it lists none and says defs_listed="0". Declaration rows have no cap of their own;
they are bounded by defs=.
* The rows are the root's first children (columnar: after the path table, before <cols>).
No new element or attribute: <def ix t sc p bodyless> as T2 printed it. The compact `def ix=`
reading and the full clauses are reworded for the declaration rows; --uses gains its own clause
(kUsesDefTableLegend: its rows name no group, so the of_def reading does not ride).
Also fixed: the compact `bodyless=N of total=` reading was read anywhere in the payload, so it
fired (107 B) on a `<def bodyless="1">` row; it is now scoped to <bodies>, the one element that
carries that count.
GATE, written first: test/verbdefrowscheck.sh, 44 arms, 24 red on the parent (57b4c3e); the
strip-identity arms pin the parent's bodies with the <def> rows removed.
RE-PINNED, each by design:
* impactsoledefcheck (3): proto.c:3 then proto.h:3 (was the one group row); (6): --callers
dropped from the "stays table-less" list.
* attribmultidefcheck: the parser reads a repeated ix as a tied-declaration row (MEMBER), not a
group; the one-group --callers arms (b_rp b_up b_q c_lost c_pick lf_stash lf_baz f_py1 j_u1)
pin one group row plus their declaration rows.
* defslookupcheck (B2, C): "one group" is now the count of distinct ix=, not "no <def> row".
* sitesatshortcheck (C): the --uses golden is compared with the <def> rows stripped.
* compactlegendcheck A-PIN callers.xml 3875 -> 4187 B, impact.xml 4844 -> 4887 B (checked
mechanically: only the rows and the clause moved); (L) loop bar 8400 -> 8800 B (measured
8789), twin 8200 -> 8700 B (measured 8644).
* test/legend_over_target.tsv: def-row 126 -> 135 B (impact, loop, mcp:impact), sole-def
230 -> 222 B (full:impact); new rows ripwire.callers/v1, ripwire.uses/v1, mcp:uses
(def-row 135), full:callers and full:uses (sole-def 222), full:uses (uses-def-table 193).
* printf_parity.manifest: callers, impact, uses, callers_json, impact_json, callers_limit.
OWED AT INTEGRATION (not editable from this lane): a changelog.d fragment; `docs/counts_snapshot.py
refresh` for the new gate (docs/TUNING.md, LIMITS.md, EVALS.md); a COMMANDS.md showcase
re-capture of --callers/--uses/--impact.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One conflict: test/printf_parity.manifest (callers_limit, help_all). The manifest is a function of the binary, so it was re-pinned from the merged build (UPDATE_GOLDEN=1), not text-merged. Moved vs lane/af1-answerfirst: callers impact uses callers_json impact_json callers_limit (V1's def rows); moved vs lane/v1-verbdef: help help_all (AF1's --page-bytes). Nothing else. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts:
# test/compactlegendcheck.sh
# test/compactlegendfix/pre_l1_full/impact.xml
# test/printf_parity.manifest
REGEN AFTER BUILD (binary-pinned files that took main's side; build, then scripts/merge_main.sh --regen-after-build BIN):
test/printf_parity.manifest
…erged V1+T3 binary The merge took T3's side of test/printf_parity.manifest (a binary-pinned file). Re-recorded with UPDATE_GOLDEN=1 bash test/printffmtparitycheck.sh on a --clean-first build of the merge: callers and callers_limit return to V1's pins (T3 did not move them), impact is new (V1's <def> rows and T3's short importer rows in one answer), help_all keeps T3's pin (V1 did not move it). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e-T3 tree T3's goldens were taken at 994f002, before T2/V1. In the batch, a one-group --impact answer also prints V1's <def> rows (and --json def_table), so arms B and E went red on the merged binary although T3's rows rebuild exactly. Re-captured with the gate's own fixture lines and the binary built from 1ddf1c29a (V1 merged, T3 not): the only difference from the old goldens is those <def> rows / def_table (legend comment and at= masked). importrowshortcheck: 39 PASS, 0 FAIL on the merged binary. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts: # README.md # docs/EVALS.md # docs/counts.tsv # present/deck5_ripwire_build.js
…ph question is answered by that verb, ROUTED
OWNER DECISION 2026-11-18 ("Accept V1, start RT1"). A bare --for whose text names a symbol
in backticks (lane P1's rule E) and asks who calls it, what it reaches or what breaks
(rule R) is answered by --callers / --impact / --uses. The verb's own answer is served,
with <route from="for" kind= verb= task= complete="not-proven" next=> as the root's
FIRST child, ahead of V1's <def> rows. complete= is never "1": nothing in a verb's bytes
proves completeness.
THE FALLBACK. Lane W1's frozen F1 (frozen_f1.py, transcribed in src/partialwarn.h) reads
the verb answer, --defs, --grep=LAST and a whole-word scan of the indexed corpus. When it
finds evidence of a miss (text-gap, macro, twin-defs, value-hop, target-floor) the route
says warn="1" why=… and the FIRST --for page rides as the root's last child, byte for byte
the opted-out answer's page 1, its own next= the second page. An absent warn= means no
evidence found, never proof. If F1 cannot run: P1's hybrid (why="kind" / "unchecked").
WHEN IT FIRES. Only on a bare --for (forUnbudgetedCall is the bare call, no --page-bytes,
no budget, XML, single root, default legend posture), and only when the verb resolves the
symbol as typed or by its last segment; anything else is the --for answer byte for byte.
--no-verb-route (the route's own next=) is the opt-out, byte-identical to the base binary.
A spelled --page-bytes never routes, so every page's next= stays a page of --for.
MCP `for` is unchanged: AF1 left MCP unpaged, and MCP clients hold the verbs as tools.
DEVELOPMENT (P1's 49 spent cells; results/RT1_ROUTE_1118): routed 21, warned 7; RT-final
41/49 complete vs --for 39/49, 0 cells losing a group, 7,533 vs 9,192 tokens/q; warn
recall 5/5, false alarms 2/16; rule E/R agree with rules.py 49/49, warn with frozen_f1.py
21/21. Confirmation on T5c's sealed set is in the readout.
GATE (written first; 22 arms red on the base): test/forverbroutecheck.sh — route fires,
does not fire (byte-identical), warn on a callback hop and a text gap, warn absent on a
plain chain, the opt-out, next= replay, legend coverage, determinism, well-formedness,
base identity (BASE_BIN=). Registered in regression.sh and the shard weights.
Re-pinned: printf_parity (help, help_all), shapingflagcheck (A) 25 -> 26 (the route reads
--max-tokens only to stand aside), fordeclrowcheck (n8) asks the --for answer with
--no-verb-route, flag counts 196 -> 197 (README x2, guide, deck x3), skills/ripwire-orient.
Legend: four <route> readings in compactlegend.h (present-only; dictv moves).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…g dropped Lane V1 (results/V1_VERBDEF_1118) measured its <def> legend at about twice its rows: --callers +5.4 KB legend vs +2.7 KB rows, --uses +7.3 KB vs +4.4 KB, pooled over 48 held-out cells. Three readings are reworded, each keeping every term a reader needs (defs= groups, a definition then its tied declarations, same ix=, bodyless=1 no body; the one-group clause's file:line, measured-from and "no row carries of_def"; --uses' "a use-site row names no group"): * compact `def ix=` row (compactlegend.h) 133 -> 111 B * full kSoleDefLegend (graphlegend.h) 220 -> 167 B (+2 separator) * full kUsesDefTableLegend, now one sentence 191 -> 151 B (+2 separator) MEASURED on V1's own cells (results/S1_SMALL_1118/fix1, V1's run/grade scripts, shared grader, matched n=48 per verb, base = V1's frozen e6671f2 binary): complete and groups unchanged on all six surfaces; bytes CLI callers 95,420 -> 94,364 (-1.11%), uses 87,631 -> 86,575 (-1.21%), impact -0.31%; MCP find_referencing_symbols 109,994 -> 107,980 (-1.83%), uses -1.15%, impact -0.31%. GATE, written first: test/verbdefrowscheck.sh (11), six arms: a byte ceiling per reading (112 / 175 / 155 B) and the terms each must still carry; four RED on e6671f2. RE-PINNED, each by design: * compactlegendcheck A-PIN callers.xml 4197 -> 4144 B, impact.xml 4903 -> 4850 B (the old kSoleDefLegend sentence replaced string for string; nothing else moved) * test/legend_over_target.tsv: def-row 135 -> 113 (6 rows), sole-def 222 -> 169 (3 rows, marker "the ONE defs= group"), uses-def-table 193 -> 153 (marker "a use-site row names no group") * printf_parity.manifest: callers, impact, uses, callers_limit * impactsoledefcheck (5): its marker is now "the ONE defs= group" Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
M2 turned the allocator-level memory ceiling on by default but could not edit src/cli.h, and left the exact replacement in its readout: the sentence that said a peak over the limit between two readings is never stopped is false for memory that goes through operator new. Replaced with M2's text (refused at operator new, exit 5, RIPWIRE_MEM_BUDGET_MB, servers unarmed; only a peak the hook cannot see still finishes and discloses). test/printf_parity.manifest: help_all re-recorded from a --clean-first build (the only label that moved; the first-screen help is unchanged). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ate declaration is a stated known gap Lane V1 (VERB-DEF-ROWS) prints a <def … bodyless="1"/> row for each declaration edit-check's identity rule ties to the definition. The rule ties the control's `void grow();` but not the template member's declaration, so --callers/--impact/--uses on the template fixture now differ from the control by exactly that row (and the present-only legend reading it brings): 3 FAIL on the merged binary, also red at V1's own tip (e6671f2), so a lane red the lane's gate run did not reach, not a composition defect. The byte-identity arm now compares the answer element with bodyless <def> rows masked on both sides, and a KNOWN-GAP arm pins the difference (control box.hpp:3 tied, template untied and silent) so tying template member declarations removes the mask rather than passing silently. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…hard weights, counts snapshot refresh - changelog.d: one fragment per change, written from each lane's readout (T2 impact one-group def, V1 verb def rows, T3 short importer rows, T4b pf= path fold, P2 tools/list trim, P3 short skill bodies, M2 memory ceiling default-on, M2 bounded head-binary build + bench timeouts). T3's fragment carries the correction its commit message (72419d3) needs: the <impact> element is 4.27 % shorter and the whole answer 3.97 %; the 3.02 % in that message is the element's saving over whole-answer bytes. The commit is not rewritten. - .github/pargates-shard-weights.json: importrowshortcheck (T3), memceilingcheck (M2) and namewalkcheck (M1) had no weight entry (the fifth gate registration, which nothing enforces); measured locally 2 s, 7 s, <1 s. - python3 docs/counts_snapshot.py refresh + docs/gatecount_build.py on the composed tree: 859 gate scripts in the loop (the set union of the seven merged branches' loops minus phantomdefcheck, which the train removed at 82ed4c8 and the 994f002-based lanes still listed). Supersedes T4b's own refresh commit 9a6d496. - skills/CONSOLIDATION.md: a dated entry for P3's body split, owed by that lane. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ined where they appear; legendcoveragecheck reaches them
Lane L1's page-legend census (results/L1_PAGELEGEND_1118) found attributes spelled `name=` in no legend
of their whole answer and on no floor. Reproduced on tiny corpora; each closed in the legend that
already defines the rest of its block (present-only, rides with the block):
* --for graph_verb from_span=/to_span= (of_span=/and_span=): forgraphverb.h said "the *_span=
attributes"; it now spells them (+19 B per graph_verb answer)
* --for shortkey found=, rung key= and q= (q= found here): kShortKeyLegend head gains
"found=0: no definition of X resolved; key= the name a rung searched, q= the Q of Q.leaf." (+89 B)
* --for d slot_fields_more= (and slot_by_more=): kForCompactSlotJoinLegend said "_more= the rest" (+25 B)
* --impact <rf by_name=>, and <rf via="name"> which only element-blind <e>/<s>/<c> rows covered: two
element-qualified kCompactAttributeReadings rows, the full dialect's fix27 reading (+56 / +129 B)
WHY THE GATE MISSED THEM, both holes closed in test/legendcoveragecheck.sh:
* Reach: no roster row produced those states. Four DEFAULT-ONLY rows on generated corpora now do
(a --for path question, an unheld qualified name, a struct with six slot fields, a cut --impact
window over a multi-definition selector partly reached by name), and arm (G-A19) fails when one
stops carrying its attribute.
* Reader: <shortkey>'s legend (before it) and the verb answer nested in <graph_verb> (legend after
its open tag) were invisible, so 22 DEFINED attributes read as gaps once reached. shortkey joins
TRAILING, a run of adjacent comments before a trailing block counts, graph_verb joins INNER, and
an attribute is judged against the head plus every block comment that contains it. Fixed-document
arms T9/T9b/T9c/T10/T10b pin it (mutation controls included).
GATE, written first: on the parent binary (G) reported 28 new default gaps (the 6 real ones + 22 reader
blindness); now (G) green, (G-A19) green, T1-T10b green. Also green: fnslotcheck, barrelshortkeycheck,
forsizetrimcheck, protoindexcheck, impactimportcheck, gvassemblycheck, graphformendpointscheck,
fordisclosurecheck, forroutecheck, legendrefcheck, compactlegendcheck, reexportcheck,
printffmtparitycheck. No re-pins; the dictionary identity (dictv) moves with the two rows.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ers says not-applicable Lane A4 (results/A4_SITU_PACK_1118, F1): --situ under a valid --legend-have= exited 0 with byte-identical stdout and an EMPTY stderr, 33 of 33, the silence the C4 rule in applyCliRefPosture exists to prevent, while --for prints a not-applicable line. The prose surfaces never reach that stage, so checkLegendHandshake now prints the same line, naming the surface, whenever legendNonXmlReadSurface() names one (--situ, --recall, --report, --mermaid, --html, --plan-lanes, --sarif, --eval*). Stdout and the exit code are untouched; a stale value is still refused first, with no extra line. GATE, written first: test/legendhavecheck.sh (S): S1 stdout byte-identical and same rc, S2 stderr says not-applicable and names the surface (--situ and --recall), S3 [N] stale still refused. S2 RED on e6671f2 (stderr empty), green now. No re-pins. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts:
# test/compactlegendcheck.sh
# test/compactlegendfix/pre_l1_full/impact.xml
# test/printf_parity.manifest
REGEN AFTER BUILD (binary-pinned files that took main's side; build, then scripts/merge_main.sh --regen-after-build BIN):
test/printf_parity.manifest
… one it replaced
- testrowruncheck arm (5) read `<t p="…"/>` document-wide, so --for's file-tail rows — which the bundle
now carries — were judged as tests_to_run rows carrying neither run= nor run_unknown=. The XML arm
reads the <tests> ELEMENT now, the same scoping the shared reader got in the commit before this.
- packtaskcheck's <far> arms are gone with the tier: there is no distance mask, so a symbol with no edge
to the key body is not demoted to a name-only row, it keeps the shape --for's own head gives it. What
was still a contract is kept and now stated as one — such a symbol gets no body and no <callers> row,
however it is ranked — plus a new arm that the removed tier has not come back. The ledger arm follows
the section order, and the tiny-budget arm reads total="1" (what qualified to be SERVED) while
asserting exactly what it always did: the element is PRESENT and says it kept nothing.
- packtaskmonotoncheck now reads TWO body denominators. `bodies_candidates` is the candidate set, which
is what its monotonicity and fixture-arming arms are about; `bodies_total` is what qualified to be
served, which is what its capped test must read. Reading the candidate count as "capped" had arm (b)
accusing a correct bundle of stranding 46-73% of its budget, because kept < candidates is permanently
true once one body is served by design and says nothing about bytes.
Pack-task gate subset: 97 of 98 green. The one left is showcasecapturecheck, whose capture records a
--legend-dict figure this lane moved; captures are regenerated, never hand-edited, and the base commit is
itself such a re-record, so it belongs to integration.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ot everything outside <tests> The first form of this fix admitted only the `<t p=…/>` rows inside a <tests> element. That looked equivalent to excluding the file-tail rows and was not: --test-gate spells tests_to_run rows that are not wrapped in a <tests> element at all, so the inclusive form silently dropped them and turned selectorchaincheck (d2) red — a gate that was green before this lane touched anything. The exclusion is now scoped to the one element that spells the other meaning: a `<t>` row inside <tail>…</tail> is a file-grain row, everything else is read as before. That cannot drop a row from a document with no tail. Verified on all ten gates that import this reader: affectedcheck, affectedfoldcheck, affectedgroupcheck, impactpartitioncheck, receiptpostcheck, rootrelemitcheck, selectorchaincheck, testgatecheck, testrowruncheck, w3fixlegendcheck — nine green, and affectedfoldcheck red for an unrelated reason (its arm (E) pins --pack-task's bytes against a pre-change capture; see the readout). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… a degrade The ASan run over --pack-task's paths printed a DEGRADED_PATH_ALERT on every --partition slice: "<bodies> did not have the expected open/close shape — the served body is dropped". Nothing was wrong. A slice's body share falls under kPackTaskSectionFloor, so packBodies is never called and bodiesXml is EMPTY; the restate then took its unrecognised-shape branch and reported a degrade for a path working as designed. A degrade alert that fires on the normal case is worse than none, because it is the alert that trains a reader to ignore the next one. The empty case is now its own branch: shown= is honestly 0, the key body rides as a pointer like every other candidate, and nothing is disclosed because nothing was lost. The alert is left for the one case it was written for — a <bodies> shape the restate genuinely cannot read. STDOUT IS BYTE-IDENTICAL. Verified on test/fixture across six invocation shapes (--token-budget=600 / 2000 / 6000, --partition=3, --json, --legend=full): 6 identical, 0 different. Only stderr changes, so the dev-set measurements taken with bin/ripwire.d1 (built from 83c55f1, one commit back) describe this code. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… pin on graphlegend.h The first FULL-suite run of the pass-3 candidate (890 gates) found three reds no lane's own targeted run had executed. None is a stale pin; each is a real defect, two of them inherited from a merge. - hashsetcheck (S): BX2's per-macro ceiling 6f023ca declared macroCeiling's `seen` as HashMap<std::uint64_t, char> and only ever asked contains() on it. The house rule is that a membership table is HashSet<K>; it is one now, with insert() in place of emplace( key, 1 ). - gateexitcheck (G2): AR1's new test/subagentprimercheck.sh reported 34 verdicts through the one-line `… && ok … || no …` spelling. G1 makes ok() safe at the definition, but G2 pins the spelling the conversion achieved, and a new gate file does not get to reintroduce it. All 34 are now the suite's wrapped form; the `case … esac` verdicts and the one reversed `&& no … || ok …` site are untouched because G2 does not name them, and the gate is green. - formatgatecheck (C): src/graphlegend.h is in scripts/formatcheck.sh's GATED set — the files that agree with .clang-format byte for byte — and f269297's DefTableLegendParts was written with aligned declarations and trailing comments, which .clang-format does not produce. Reformatted with clang-format 22.1.8 (WANT_MAJOR=22); only that struct's four lines moved. Green at this commit: gateexitcheck, formatgatecheck, hashsetcheck, and the behaviour gates around the two touched symbols — verbfixcheck (its [F2C] arms gate the very ceiling macroCeiling implements), rustmacrocallcheck, attribmultidefcheck, verbdefrowscheck, mcpclidiffcheck, legendcoveragecheck, capdisclosemapcheck. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…uite found
The 98-gate pack-task subset was run before this lane's last few commits and reported 97 of 98 green. The
full suite (890 gates, 869 pass, 3 skip, 18 fail) found eleven more, and these are the ones that cannot
change an answer byte:
- changelogfragcheck / mcpmanifestcheck: my changelog.d fragment's first line has to be
`### <Section> — <title>`. It was a bare title, so scripts/changelog_assemble.py refused the whole
directory and two gates reported it.
- fixedbufsweep (S2/S4): the committed fixed-buffer registry is a per-buffer COUNT plus prose, and four
rows had rotted under this lane: `b` 10 -> 11 (the ledger's bodies line), `bodylessAttr` 2 -> 1 (the
bare-wrapper placeholder branch is gone — a bundle with no body CANDIDATE has no bodyless owner to
count), `tag` 2 -> 1 and 112 -> 144 B (the render-failed branch is the survivor and gained a capped=
field), and the `farAttr` row is deleted with the <far> tier it described. serialize.h's
xmlCommentText CALL-SITES list goes 17 -> 16: the W2-K restated body-omission marker was that site, and
it went with restatePackTaskBodiesWrapper.
- limitstablecheck / readmedriftcheck: docs/LIMITS.md regenerated (446 caps, 10 parameters), which this
lane's new constants had made stale. Generated, not hand-edited.
- gateexitcheck (G1/G2): my three gates spelled ok() without the accumulator every other gate has — a
failed write of a PASS line turned a red arm green — and reported verdicts through the one-line
`cond && ok … || no …` shape, which G2 forbids because a failed write makes it print FAIL for an arm
that passed. All 23 verdict sites across the three are if/else now, and ok() is the house pair.
Not in this commit, and not hygiene: bodydialectcheck, bundleidcheck, callsrankordercheck,
compactlegendcheck and forsoftbudgetcheck are five REAL output-contract defects the suite found, each of
which changes answer bytes and therefore invalidates this lane's 596-answer measurement. They are reported
in the readout for a decision rather than fixed here, because the right scope for them is one iteration
together with the LocBench tail-starvation finding — see READOUT §5a and §6b.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Both were red for the same reason and neither was a disagreement about behaviour: they pin bytes this lane
changed on purpose. Regenerated, never hand-edited.
- test/affectedfoldfix/golden/{big,brace,odd,small}.pack-task.xml, through the gate's own capture.sh.
The FIRST attempt at this (earlier in the lane, reverted) passed capture.sh a RELATIVE binary path;
capture.sh cd's into the fixture directory before running it, so the binary was not found, every
pack-task output was EMPTY, and those empty files went over the committed goldens. This time the path is
absolute and nothing is copied until each candidate is proven non-empty, over 1,000 B, xmllint-clean and
carrying schema="ripwire.pack-task/v1". The four grew 8,216 -> 10,841, 7,639 -> 9,554, 7,193 -> 9,100 and
3,374 -> 3,632 B, which is the file tail arriving.
- docs/captures/COMMANDS_showcase_2026-11-19.md through test/showcase_capture.py (270 commands, rc=0),
then docs/COMMANDS.md rebuilt FROM that fresh capture — in that order, because COMMANDS.md lifts its
samples from the newest capture and rebuilding first would have carried the stale legend-dict figure
onward.
asan/ was rebuilt to the same commit as build/, which is what mcpreloadasancheck refuses without (it
compares two binaries' answers and will not compare across source trees).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…say so
bodydialectcheck found it and it is the real thing: capped= on <bodies> read only `served < servedTotal`,
so a body that WAS emitted and then cut short reported capped="0". MEASURED over this lane's 596 dev
answers: 194 of them — 33% — carried `<b … truncated="1">` under `<bodies capped="0">`. A truncation the
document does not admit to is the same defect class as a false cap with the sign flipped, and it is the
reading truncvocabcheck already spells for this element ("rows all present but one is CUT: capped=1 with
shown == total").
capped= is now `served < servedTotal || servedCut`, the ledger's "(capped)" marker follows the same rule,
and the legend says it. The fix is BYTE-LENGTH NEUTRAL — `capped="0"` and `capped="1"` are the same width —
so the byte medians and the completeness counts in the readout's §5 are unaffected; one character changes on
194 answers.
My own honesty sweep (pb1_honesty.py H1) missed this, and the way it missed is worth recording: it allowed
capped="1" with shown == total when a row carried truncated="1", and never checked the converse. It had the
right concept and applied it in one direction only. H1 now reads both ways.
Also: PackBodiesTag is initialised with DESIGNATED initialisers. Adding servedCut mid-struct silently
re-slotted a positional aggregate — countBodylessCandidates() landed in servedCut and `false` in bodyless —
which compiled and would have inverted two disclosures.
bundleidcheck: its ledger parse is `grep -o 'ceiling [0-9]*' | head -1`, which matches "ceiling " with ZERO
digits. This lane's legend gained the phrase "the ceiling cut", so head -1 took the prose and every budget
reported "the ledger is unreadable". A gate reading a NUMBER must require a digit; tightened at its three
sites.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…s budget's forsoftbudgetcheck (P2) found that `--pack-task --token-budget=6000` served different sections from the DEFAULT --pack-task, whose budget is also 6000. Located: packTaskHeadRules was handed `in.softTarget ? budgetTokens : 0`, and forHeadRoundRobinOn() reads that ceiling and returns false at 6000 — so the default (0) emitted `files_first=` and the explicit twin did not. MEASURED on this repo: `<sigs … docs_dropped="22" files_first="15">` against `<sigs … docs_dropped="22">`. Two things were wrong with that, not one. It contradicts the verb's documented contract — "an explicit 6000 serves the default's sections byte for byte", which is kPackTaskHardBudgetSentence's own comment — and it silently disabled, for EVERY explicit budget, the head round robin by file that PT1 measured as one of D1's two locating mechanisms (4 of the gold files it recovers are head rows only through it). The fix passes 0 in both paths: the default's verdicts are the ones both must share, and the default has no caller budget to read. The byte budget still cuts the section — these predicates only decide whether the head spends bytes on e=, the docs reorder and the round robin. VERIFIED: the default answer on this repo is BYTE-IDENTICAL before and after, so the 596-answer dev measurement describes this code unchanged. Only explicit non-default budgets move, and they move toward the default. packlocatecheck, packkeybodycheck, bundleidcheck and truncvocabcheck re-run green. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… any section spends
PT1's design sentence for the file tail was "it takes its own share PLUS everything the sections before it
leave". The first D1 build implemented only the second half. The tail was the leftover with no floor, while
the head was allowed up to kPackTaskHeadRowDropPct (60%) of the post-header budget rather than its own 45%
share, and monotoneRoll charges a cut section its whole granted share. On a corpus with a long task echo and
long signatures the leftover was nothing.
MEASURED over LocBench's 303 tasks: the tail declared a 20% share and received 0.6% — a median shown="0" of
total="512" files — and on 82 of the 84 tasks the bundle failed to complete, the gold file it needed was in
that cut tail. Completion is counted in FILES and the tail is the section that names files, so a starved
tail is a bug and not a tuning shortfall. The cross-set picture says the same thing: the tail is fed on every
set that HIT its bar (SUB1 46 of 716 files, SFML 51 of 232, Multi-SWE 18 of 166) and starved to the bare
83 B shell on the one that missed.
The reserve is withheld from the spendable pool up front and handed back to the tail at the end, so the tail
still absorbs every leftover as well; quotaBase stays the full post-header budget so the other shares keep
their own percentages of it (45+20+8+2+5 = 80 = the pool left once the tail's 20 is reserved, so nothing is
over-committed by construction). A --partition slice reserves nothing: it carries no tail. The static_assert
now requires the head's bound to leave the key body AND THE TAIL their shares.
MEASURED on a 240-file fixture with long signatures and a 1,200-char task — LocBench's shape: the tail goes
from shown="0" / 46 B to shown="7" / 1,693 B. On this repo with a 1,200-char task: 83 B to 2,028 B.
test/packtailsharecheck.sh proves it, and is RED-FIRST against the pre-fix binary (arm (R) asserts that, so a
gate that stops exercising the defect says so instead of passing quietly). Its first fixture used 12 short
parameters and the pre-fix binary produced a byte-identical tail — arm (R) caught that the gate was vacuous,
which is why the fixture now carries 40 long-named parameters. Listed in test/regression.sh.
Also in this round, each closing a gate the full suite had found:
- selfcheckcheck: --pack-task's five sink-less DISCLOSE( msg ) sites took the PackTaskRenderFaults sink
(HeadShapeUnknown / BodiesShapeUnknown). The ratchet counts the sites that are SILENT IN RELEASE and its
stated target is 0, so the pin was NOT bumped to absorb the three this lane added — converting them took
the count 51 -> 46, below the old pin of 48, and the pin follows it DOWN.
- compactlegendcheck: the compact legend was re-tightened (every reading kept, the wording around them cut)
and the two clauses D1's answer needs that the 990 B target predates — the <bn> pointer row and the file
tail's rows — are registered in test/legend_over_target.tsv as present-only, with their reasons.
The kept ledger no longer contains a '>': it rides as a DATA comment that consumers slice with
`<!-- ledger: [^>]*-->`, so "next= on <tail>" made the whole disclosure read as absent (FX4).
P5 compares the tail fact by section and by whether it still discloses a cut, not verbatim: the tail is
filled from what the DELIVERED document leaves and the compact legend is smaller, so the same budget
legitimately buys the compact answer more tail rows (MEASURED: "kept 0 of 142" full, "kept 6 of 142"
default). Holding that to a verbatim match would force the tail to be cut against a fixed legend
allowance instead of the bytes actually delivered, which is what makes the per-answer ceiling hold.
- callsrankordercheck and bodydialectcheck: their fixtures name the served body in backticks, because
<callers> and <calls> are the KEY BODY's now; bodydialectcheck's omission-parity arm compares the
--expand SELECTOR rather than a name recovered from it (a candidate can be a markdown heading whose own
name contains ": "), its roster drops the removed <far> tier, and its (C) arithmetic arm gained the
"rows all present but one is CUT" allowance its own (A) arm demands — without it the two arms contradict
each other and no binary can satisfy both.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ut of packTaskBundleText
Owner ruling, 2026-11-19, on this lane's sizing: cognitive complexity 226 -> 348 while cyclomatic went only
129 -> 156, so +27 of the +122 was new logic and the rest was nesting weight. BEHAVIOUR DOES NOT MOVE — see
the byte-identity proof below.
Two cohesive blocks come out, and deliberately NOT the whole body under a new name: a rename would have
moved the finding onto a new symbol instead of fixing it, which is the flaw this lane reported in its own
cheaper option.
- packTaskJsonTail( j, PackTaskJsonParts ) — 170 lines. The JSON dialect re-serializes the SAME section
decisions the XML above already made, so it needs those decisions and nothing else. Its inputs are a
PARTS RECORD following verbs_quality.h's QualityDeltaLegendParts: a flat record of already-computed
values, const references for the containers, one comment per field, no std::function and no owning
copies — read once, inside one call, every source outliving it. 33 fields, two of them deliberately
non-const references because the callee takes their address (forClone's lens pointer, testPartition).
- packTaskSettleRoot( whole, headerStr, buildHeader, lr, in, budgetTokens ) — 106 lines, the ceiling
ladder and the priced root attributes. SIX inputs and no record at all: the block reads almost nothing
from its surroundings, and the header BUILDER arrives as a template parameter because
climbCeilingLadderBy already takes it as a callable, so nothing here needs an allocation.
MEASURED, on this repo:
symbol cx ccx
packTaskBundleText 129 -> 113 226 -> 215 (base -> now: BELOW its own baseline on both)
packTaskJsonTail new: 34 new: 48
packTaskSettleRoot new: 13 new: 17
--quality-delta against the share-floor HEAD: regressions=4, all four new-symbol rows, gating=0. The
complaint was not relocated: the pre-existing symbol improved and neither new symbol is a monster.
BYTE IDENTITY, which is the only thing that makes a refactor safe. 240 invocations — three corpora
(test/fixture, src/, the repo) x five tasks including a 320-char and a 1,200-char one x {default, --json,
--legend=full, six token budgets x {XML, JSON}, --partition=3} — all 240 byte-identical to
bin/ripwire.sf, the frozen share-floor binary. The whole-dev-population diff over all 596 answers is the
commit that follows.
The first attempt was NOT identical, and that is the point of the proof: a word-boundary rename rewrote the
identifier inside a STRING LITERAL, so every --json answer carried `{"p.task":` and differed by exactly 4
bytes. 105 of 240 invocations caught it. Gates would not have: none of them reads the JSON root's first key.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The share floor, the compact-legend trim and the refactor landed after the captures were last regenerated at ffc6b75, so three absolute pins were red at the tip. None is a disagreement about behaviour; each pins bytes this lane changed on purpose. All three regenerated through their own generator, never hand-edited. - test/affectedfoldfix/golden/{big,brace,odd,small}.pack-task.xml, through the gate's own capture.sh with an ABSOLUTE binary path (capture.sh cd's into the fixture dir; a relative path silently writes empty goldens, which is how the first attempt in this lane damaged them). Each candidate was proven over 2,000 B and xmllint-clean before it was copied. The four shrank 10,841 -> 10,793, 9,554 -> 9,506, 9,100 -> 9,052 and 3,632 -> 3,584 B: 48 B each, and the token-level diff shows the whole of it is the compact legend's shorter prose ("the ONE key body served" -> "the ONE key body", and three more clauses), with est_tokens following it down 1,528 -> 1,508. No row, no attribute and no body moved. The other 29 files capture.sh writes are untouched and still byte-identical to the base commit. - test/printf_parity.manifest, with UPDATE_GOLDEN_EXPECT=pack_task so the re-pin had to name the one label it was allowed to move: moved={pack_task}, 41 unchanged. - docs/captures/COMMANDS_showcase_2026-11-19.md through test/showcase_capture.py (270 commands), then docs/COMMANDS.md rebuilt FROM that fresh capture, in that order: COMMANDS.md lifts its samples from the newest capture, so rebuilding first would carry the stale legend-dict figure onward. The recorded --legend-dict figure now reads dictv=98a9457a5d5eb284 entries=1287, which is what the live binary prints. Both build trees were rebuilt --clean-first at this tip first, because build/ was stamped e8f0dd5+dirty and asan/ 6e22319, five commits behind: versioncheck, versionlabelcheck, g1freshcheck, mcpreloadasancheck and cachefuzzcheck were all red on binary staleness rather than on anything in the source. The rebuilt build/ripwire is byte-identical to the frozen bin/ripwire.rf across all 240 --pack-task invocations of the byte-identity spread, so the re-stamped binary is the measured one. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The entry described D1's shape and the token conversion but said nothing about the three behaviours that
landed after it was written, and a changelog that records only the first half of a change is a defect of the
same kind as a readout that records only wins:
- the file tail takes its declared share as a FLOOR, with the measurement that made it necessary (the tail
declared 20% and received 0.6% over LocBench's 303 tasks; the gold file was in the cut tail on 82 of the
84 tasks that did not complete) and the result (329 -> 420 of 596 bundles naming every file needed), plus
the statement that the budget did not change.
- <bodies capped=> now admits a served body that was truncated: 194 of 596 answers published capped="0"
over a <b truncated="1"> row.
- --pack-task --token-budget=6000 now serves what the default serves; the two differed by a section's
reading because the explicit form passed its own budget into the head's lens rules.
Also fixed a spelling error in the entry itself: the pointer row is <bn … next="--expand=…"/>, not <b>, and
added test/packtailsharecheck.sh to the gate list.
The 891-gate suite ran at 94e149e, one commit before this one; this commit touches a single changelog
fragment and no source, so the 15 gates that read changelog.d were re-run at this tip instead of the suite:
binoverride, changelogfrag, deck, docdrift, docdemote, forliteral, forunused, mcpmanifest, mdembed,
prcontext, qschemetrip, qualitykinds, quality, queuehop, selfcheck — all green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…re the code The gate before the code it measures (non-negotiable 1). Twelve arms pin the DELIVERY UNIT of the one capped-body block --for and --pack-task are to share, and nothing about ranking: whole body at <=120 lines of indexed span (e= - l= + 1, the denominator SR1 measured and the only line count the index knows), HEAD-ONLY past it with the head's size stated, the cut disclosed through truncated=/lines=/next=, its CAUSE readable through a new cap_lines= so a line cap is not mistaken for the byte cut whose registered reading says "cut at the byte budget", a zero that says what was scanned, one block rather than six, and no anchor window. The no-anchor arm (C6) is the one that keeps SB1's refutation enforced: a +-30 window anchored on the query's best term holds the edit in 20 of 63 large bodies (31.7%), falling to 19.0%, so the unit is head-only and a later lane may not reintroduce the window silently. cap_lines= was checked free before it was introduced. head_only= was REJECTED: src/abicheck.h already owns it on <ref> for --abi, where "head" means git HEAD. RED on the frozen integration binary (results/INT1_070_1119/bin/ripwire-070cand-p3), on a fixture, never by breaking a real document: C2 C3 C4 C5 C6 C7 C8 C9 C10 fail on the rule's absence; C0 (non-vacuity, including a span self-check against the binary's own index), C1 (bodies that already fitted), C11 (determinism) and C12 (xmllint) pass as controls. No arm passes for the wrong reason -- an earlier draft of C8 counted <b> rows anywhere and so passed on three bodies arriving through three different shells, and an earlier C9 compared two all-absent units; both now carry non-vacuity guards. Registered in test/regression.sh in this commit (manifestcheck green, 891 gates). The published count snapshot in docs/counts.tsv is left alone: it is a build product and regenerating it serialises every lane that adds a gate. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> (cherry picked from commit 3799ff4247afe85c751f0b7aff13ba0f82be0439)
TWO KNOBS, BOTH DERIVED RATHER THAN PICKED. 1. kCappedBodyLines = 120, applied as min( span, 120 ) -- monotone, no discontinuity. The owner ruled this shape over SR1's "whole <=120, else a 15-line head": that shape saves more (CB1 measured -42.3% of candidate body lines against this one's -21.7%) but an answer whose size jumps 8x on a property the caller cannot see is the wrong trade, and it is worst in the fan-out shape where N subagents each budget alone with nothing to normalise them. 2. kKeyBodyMax 2 -> 5. GA1 DEV's gold functions per task run median 1, q75 3, p90 5, max 12, so a cap of 2 could satisfy 71.9% of tasks and 5 satisfies 94.4%. The guard sits at p90 and the byte budget is what actually bounds the block. The premise offered for lifting it -- "the median task has 3 gold files" -- does NOT hold on GA1, whose DEV median is 1 file and 1 function; only 28.1% of tasks need 3+ functions. The raise helps a real minority, and the readout says so rather than inheriting the stronger claim. WHERE IT APPLIES, and why not everywhere. Three surfaces serve "this candidate's definition body" and all three take the cap: <bodies of="key">, <bodies pick="name|member">, and --pack-task's. The route, hops and launch groups keep bodyLineCap=0 -- their unit is a path through several definitions, not one candidate's body. Which of the three fires depends on the query, so a rule bound to one of them would be a rule a caller cannot predict; the gate therefore asserts the UNIT, not the shell. DISCLOSURE: cap_lines=. truncated=/lines=/next= already say how much was cut, but compactlegend.h registers truncated= as "cut at the byte budget", so reusing it alone for a LINE cap would make a registered reading false (non-negotiable 3). cap_lines= names the cap that cut; a byte cut carries none, so the two causes are distinguishable from the row alone. head_only= was rejected: src/abicheck.h owns it on <ref> for --abi, where "head" means git HEAD. BYTE IDENTITY, by construction and then measured. bodyLineCap defaults to 0 through packBodies, renderBodyElement and routeBodiesXml, so every caller that does not opt in cannot reach capBodyAtLines. 17 non-vacuous verb arms on htop are byte-identical (map, hotspots, lint, deps, clones, dead-code, test-gate, impact, uses, callers, around, expand, path, exemplar, lego, grep, pr-context); an 18th (situ) compared two empty outputs and is reported as vacuous, not as a pass. hopbodycheck (P4) independently shows a 154-line HOPS body still cut at lines="1-46/154" with no cap_lines=, which is the uncapped groups proved by output rather than by reading the defaults. GATES. test/cappedbodycheck.sh was red on this tree before this commit (9 arms) and is green after, with C7 and C8 printing as DEFER: the owner put the six-shell consolidation out of scope, so those two bars live in PREREG_CB1_1119_v2.md D1/D2 rather than being deleted or left as an unclearable red. An earlier C2 asserted a marker's absence and failed on correct output -- with a 121-line span and a 120-line cap exactly one line is cut and it is the closing brace -- so C2 now asserts the arithmetic. test/forkeybodycheck.sh K6 is re-worded in THIS commit as the knob move requires, and a new K6b uses six names so the cap of 5 is exercised rather than merely outnumbered; K6 alone would pass on any cap >= 3. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> (cherry picked from commit f481c6b0dcb7207ddd45d1eb136e247778a9d780)
…uilt
MECHANISM 2 IS NOT COMMITTED. Only its gate is. The implementation was written, measured, and reverted.
WHAT RC1 MEASURED, and why the failure is worse than its frequency. 8 of 69 release first calls exceed the
client's 30,000-character limit -- 11.6% against 1 of 234 on dev, about 27x, Fisher p = 3.8e-05 -- and it is
a routed-verb problem: 7 of 14 routed --impact, 1 of 9 uses, 0 of 39 unrouted --for. All 8 crossers carry a
--for page; none of the 10 non-carriers crosses. When an answer crosses, the client replaces it with a
~2.3 KB envelope whose 2 KB preview lands INSIDE THE LEGEND on all 8, so the agent gets ZERO answer rows and
complete=/has_more= go with them; in 5 stored persisted cases 0 of 5 agents reopened the saved file.
WHAT I BUILT AND THEN REVERTED. A descending spec ladder on the carried page's render took it from 14,531 B
to 8,292 B, -43%, with the page still present and page1="1" intact, so EX1's +65K SE/question survives.
Two measurements killed it:
1. The bound I pre-registered was on bytes the mechanism cannot move. The page is 8,292 B of which
3,034 B (36.6%) is its own legend, and the parent routed answer carries NO legend -- so the page's
legend is the document's only definition of the attributes its rows emit and legendcoveragecheck would
fail without it. The ladder bottoms out at a spec of 800 and the page still emits 8,231-8,292 B. A total
bound of 8,000 B is unreachable at any spec.
2. It shrinks the page SILENTLY. The carried page's root carries no page=/pages=/withheld_bytes= at all,
before or after, and setting both forPageBytes and pageBytesSpec on the reflow does not produce them --
the unrouted paged call does emit them, so they are stripped or never attached in the routed assembly.
An agent would get 43% fewer rows with nothing saying rows were withheld and no next= to recover them.
That is the silent cut non-negotiable 3 exists to forbid, so the mechanism does not ship to clear a
byte bound.
THE GATE IS THE DELIVERABLE, and its (R3) is the arm that stopped the ship. (R3) first PASSED on the silent
cut, reading an absent set of paging attributes as "nothing was cut" -- the shape the coordinator warned
about, where a must-not-print check passes on unfixed code by construction. It now compares against the base
binary and FAILS on a page that is smaller without disclosing. (R0) is the non-vacuity guard and it earned
its place: the route fires only on a BARE --for=Q, so my first five attempts to reproduce a carried page
passed --legend=compact and silently removed the entire subject of the gate. (R2) pins shrink-never-drop,
(R4)/(R5) are byte-identity controls against RIPWIRE_BASE_BIN. R1/R1b/R3 print as DEFER with the blocker and
their bars live in PREREG_CB1_1119_v3.md G2/G3/G6/G7.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
(cherry picked from commit 29651e1b37d56f5ce576c8603c31225294872593)
…t headers The generated half of the lane, rebuilt here rather than carried across. docs/LIMITS.md is REGENERATED from this branch's own constants (§8: generated files are regenerated, never text-merged) and the count proves why that matters -- 446 caps here against 447 on lane/cb1-code, because that branch also carries p3's lineage. Cherry-picking the file would have imported a cap count this tree cannot produce. The two changelog fragment headers take the corrected form `### <Section> — <title>`; the cherry-picked commits carry the pre-fix `### Changed`, which changelogfragcheck and mcpmanifestcheck both reject. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…cords a lost line TWO REAL REGRESSIONS THIS LANE CAUSED, found by running the suite against PB1's TIP ITSELF rather than against my merge commit. Both are in gate code, not in the product. 1. forbodyslotcheck (MC5) read "no <b n=render_screen>" on a document that plainly contains it. Its bodies() parser did not strip comments, and the legend DEFINES `<b …>` rows by printing their own tag text -- so `<b ` appeared twice (once in a comment, once as the real row) while `</b>` appeared once. The regex matched the LEGEND's `<b `, then ran `(.*?)</b>` to the real row's close, swallowing that row's open tag into the body group: attributes came back EMPTY and n came back None while the CDATA still looked right. This lane's new `cap_lines` legend entry, which begins `<b `, is what pushed the comment count from zero to one and exposed it. The gate already strips comments for its ROW parser (strip_slot, same file), so this is its own idiom applied where it was missing. Same class as the miscount OR1 recorded, where a legend's own <bodies> was read as an answer's. THE PRODUCT WAS NEVER WRONG, and that is measured, not assumed: the CLI's and MCP's <b> rows are byte-identical (lines="1-47/176" truncated="1" next= cap_lines="120" on both) and their CDATA is byte-identical at 4,876 bytes / 47 lines. I had reported an MCP/CLI parity break, then retracted it on a bad reference, and the parity was intact throughout. 2. gateexitcheck (G1): ok() in this lane's two new gates printed a PASS line without recording a FAILED WRITE, so a lost line would be silent and the gate would still exit 0. Both now use the house form, and defer() got the same protection. HOW BOTH WERE FOUND, because the method is the point. My reference binary was built at my own MERGE commit, which carries PB1's tip AND p3's lineage -- two differences from the baseline, so "fails on the reference too" could not tell "pre-existing" from "mine". Building PB1's TIP ITSELF settles it: its legend dictionary is dictv=98a9457a5d5eb284 entries=1287, exactly what the committed capture records, and forbodyslotcheck, gateexitcheck and showcasecapturecheck all PASS there. Three of the clean branch's five suite failures were mine, not one. STILL RED, AND DELIBERATELY: showcasecapturecheck. Mine, and the arithmetic is unambiguous -- the capture records entries=1287, PB1's tip is 1287, this branch is 1288 because `cap_lines=` is a new dictionary entry. The gate says "regenerate the capture, do not hand-edit the number". I am NOT regenerating it in a lane: test/showcase_capture.py writes docs/captures/COMMANDS_showcase_<TODAY>.md, and docs_commands_build.py's newest_capture() takes sorted(glob)[-1] -- so a lane-side regeneration plants a capture that every later COMMANDS.md lifts from, which is the propagate-a-pin trap. It belongs to the integration capture re-record, with this commit naming the exact cause. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`promoteGuarded` walked the post-tail sections ONCE, in emitter order, inside a
finite guard headroom, so a section won a promotion slot by its POSITION rather
than its value. That is fine while every promotable name is signature-shaped, and
it breaks the moment a body-bearing section joins the same list: in verbs_for.h's
emission order the body sections come before `lit`, so a body consumed the
headroom `lit` had been using and `lit` was deferred to a later page, silently
un-promoting a section AF2's measured gains were built on (legendseencheck went
green->red; measured by lane R1).
AF2's frozen `kPromotable` is left exactly as frozen and a second, lower-priority
class is added beside it, `kPromotableBody` = { "fold", "bodies" }. The walk
becomes two class-ordered sweeps over the same list, in the same emitter order,
against the same single headroom accumulator:
pass 1 promote every kPromotable section that fits -- the parent's own walk
pass 2 promote every kPromotableBody section that fits in what pass 1 left
and the emitted order is [already ahead of tail] [promoted RES] [promoted BODY]
<tail> [not promoted]. Pass 1 reads the same list in the same order against the
same stop with `used` starting at 0, and a section of no promotable class never
consumed headroom before either, so pass 1 promotes exactly what AF2's G promotes;
because pass 2's sections are emitted BEHIND pass 1's, their byte offsets do not
move either. Crowding-out is therefore removed BY CONSTRUCTION -- there is no
input for which a body can take a RES section's slot -- and when pass 1 exhausts
the headroom nothing moves and the answer is the parent's byte for byte.
Same guard, same stop (kAnswerPageBytesDefault), same emitter order within a
class, tail rows 1..kGuardTailRows still never move. The one class sweep is
`promoteClassSweep`, a named file-scope helper rather than a lambda, because the
first version pushed promoteGuarded's complexity from 11 to 18 and gated
--quality-delta; the refactor is byte-identical on 89 of 89 measured pages.
MEASURED on a fresh orientation gold set (89 tasks, 11 repos, 11 languages, one
worktree per task at its own base commit), three arms with identical argv:
gold FILE named on page 1 unchanged in every cell and every repo (0 down)
edit FUNCTION named +2 to +4 cells, 0 down (definition-independent)
edit-function BODY delivered 9 -> 20 cells of 89, 14 up, 0 down
page-1 bytes delta median 0, max +679 B, 0 cells over +1,500 B
whole answer never grows; population sheds 481 B / 2,145 B
byte-identical no-op 56 of 89 cells
and the mechanism bar: RES_ahead and deferred= identical to the parent in 89 of
89 cells, against 3 cells where widening kPromotable itself displaced `lit`. The
instrument is armed -- it fails on that build -- so its green carries information.
What this does NOT fix, stated because no ordering rule can: the guard reserves
tail rows 1..24 and nothing more, so on a cell whose tail overflows the page any
promoted body still pushes the DEEPEST tail rows to a later page. That is not a
priority question; it is a zero-sum trade between more file pointers and one body,
and it is why 8 of the 9 file-rung losses lane R1 measured on a map/orient corpus
survive this change while the 1 caused by crowding-out does not. Settling it needs
a cue read from the question, not an ordering (owner ruling 86, its own lane).
Gates: foranswerfirstcheck gains G6/G6n/G7 -- the RES section is still promoted
AND delivered on page 1 with its block-local legend, and no RES section was
displaced by a promotion. G7 asserts an absence, so it was proved red against a
scratch build that widens kPromotable instead, read off that binary's generated
output. The AF2 arm's RES tuple is restored to AF2's ten with a separate body
class beside it. legendseencheck is green again (3 block-local lit legends, the
parent's own value). forkeybodycheck (K3) is RED and left red: a promoted
<bodies pick="name"> makes KB1 skip a body already served on the page, so its key
group carries one named definition instead of two. No body is lost from any
answer; that is KB1's contract and a separate change (owner ruling 87).
G1: asan tree built --clean-first, 8 runs rc 0 with zero stderr, seven of them
targeted at the changed path because the default map never enters answerFirst.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ion/070-1120
# Conflicts:
# docs/COMMANDS.md
# docs/captures/COMMANDS_showcase_2026-11-19.md
# test/printf_parity.manifest
REGEN AFTER BUILD (binary-pinned files that took main's side; build, then scripts/merge_main.sh --regen-after-build BIN):
docs/COMMANDS.md
docs/captures/COMMANDS_showcase_2026-11-19.md
test/printf_parity.manifest
…ion/070-1120
# Conflicts:
# README.md
# docs/COMMANDS.md
# docs/REFERENCE_GUIDE.md
# docs/captures/COMMANDS_showcase_2026-11-19.md
# present/deck5_ripwire_build.js
# skills/ripwire-mcp/SKILL.md
# src/cli.h
# src/defattrib.h
# src/ingest_cache.h
# src/mcpverbs.h
# src/quality.h
# test/compactlegendfix/pre_l1_full/callers.xml
# test/legendcoveragecheck.sh
# test/printf_parity.manifest
# test/qschemetrip.hash
# test/qschemetripcheck.sh
REGEN AFTER BUILD (binary-pinned files that took main's side; build, then scripts/merge_main.sh --regen-after-build BIN):
docs/COMMANDS.md
docs/captures/COMMANDS_showcase_2026-11-19.md
test/printf_parity.manifest
…nd pair, COMMANDS.md)
scripts/merge_main.sh --regen-after-build, from: ripwire 0.6.10-dev.160+gc17dc32b6 (dev, AppleClang 21.0.0.21000101, emit=std::print, built_from=c17dc32b6)
changelog.d/proven-site-keeps-namesakes.md
docs/COMMANDS.md
docs/captures/COMMANDS_showcase_2026-11-19.md
test/printf_parity.manifest
parity labels re-pinned: callers callers_limit help help_all impact uses
The hash is a build product of this tree, so neither merged side's value could be right once kParserVer and kIngestParserVerMirror move to 332 (above main's 331 and the 0.7.0 stack's 317). Re-pinned once with UPDATE_GOLDEN=1 on the merged build c17dc32; the log entry names the hash and both superseded values. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
Important Review skippedToo many files! This PR contains 319 files, which is 19 over the limit of 300. To get a review, reduce the PR to 300 files or fewer by splitting it into smaller PRs or changing its base branch. Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available.
|
…n compile them `windows (cl, ninja)` failed on PR #710 with two errors and nothing else: src/forroute.h(196): error C2326: 'hasCallWord::<lambda_1>::operator ()': function cannot access 'kSuffixes' src/forroute.h(393): error C2326: 'pathNounThenPrep::<lambda_1>::()::<lambda_1>::operator ()': function cannot access 'kPreps' Both tables were function-local `constexpr` arrays odr-used from INSIDE a lambda (one nesting level for kSuffixes, two for kPreps). MSVC refuses a lambda's access to an enclosing function's automatic constexpr table; clang-cl, AppleClang and gcc all accept it, which is why `windows (clang-cl, ninja)` passed on the same tip and no other job in the matrix saw this. Giving both tables static storage duration removes the question rather than working around it: a variable with static storage duration is never captured, so no implicit capture is required of any compiler. This is also the house form — `src/` already has 286 in-function `static constexpr` tables, including the sibling `kNouns` in src/forfilepage.h. Not a behaviour change: the cues, their order and the rules that read them are untouched. The remaining 19 function-local `constexpr` tables in `src/` are passed as the ARGUMENT to the algorithm and so are evaluated in the enclosing function's scope, never accessed by a lambda; `cl` compiled all 19 in the run that rejected these two. UNVERIFIED LOCALLY: this machine has no MSVC, and AppleClang accepted the code both before and after, so the local build cannot distinguish the two states. CI's `windows (cl, ninja)` job is the only instrument for this fix. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ports The `qd-range` job fails on PR #710 with exactly the two gating rows IN3 measured locally (its §6.2 ref-pair run: regressions=111 minor=75 gating=2 — CI reproduced all three figures): complexity src/serialize.h::rw::packBodies (src/serialize.h:8199, 101 -> 115) complexity xml_rows (test/testrowpaths.py:142, 26 -> 33) Both are inherited by this integration, not introduced by it, and I re-proved it independently rather than carrying IN3's claim: - `git diff --name-only 0295652 HEAD -- test/testrowpaths.py` is EMPTY, so that file is byte-identical to lane/r1b-priority. - no hunk of src/serialize.h between 0295652 and this tip touches packBodies: `git diff -U0` hunk headers jump 3870 -> 10033, and packBodies is at 8199. Neither row was acked on any contributing lane because each lane's own --quality-delta range kept it under the gating bar; they gate only at the aggregate ee8d1a5..HEAD range qd-range measures, which sums 144 non-merge commits. So this is a range-aggregation effect, not a regression this PR performs. Acked with --ack-only so the integration's CI reports the integration's own defects rather than two pre-existing ones. Decomposing packBodies belongs to the lane that owns it, not to a merge resolution. Verified after the write: `--quality-delta=ee8d1a50e..HEAD` reports baseline="ref-pair" gating="0" rc=0 (acked 15 -> 19; the four rows are packBodies' complexity/params/verbosity and xml_rows' complexity). REVERSIBLE AND REVIEWABLE: this is a judgement, not a defect fix. Dropping it is deleting these ledger rows. Note the tool's documented heal-forward overwrote the prior reasons on three packBodies rows (WAVE-2 close, lane FE-B, V1 harvest); they are recoverable verbatim from this commit's parent. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…so a merge-heavy tip can regenerate chooseInWindow refused to write the whole 270-command capture because the --in=DIR case had no usable --since window. Measured on integration/070-1120: the available change windows touch 1 src/ file (HEAD~2..HEAD~4) or 53 (HEAD~5 and every anchor beyond it, the merge's own two parents included), while the case needs 4..22 — the merge of main contributes 53 files by itself, so the cliff is one commit wide and no anchor search can cross it. The window was not what failed. In the wide window the tool's answer is COMPLETE: the scoped <recent scope="src"> page carries capped="1" has_more="1" next_offset= next=, and the <symbols stubbed=> stub follows it. Only the capture's own 30-display-line cut hid them, behind a global <recent> block the tool caps at 40 rows. So the cut now keeps the trailing ELEMENT a caption promises, exactly as it already keeps a trailing LEGEND (trailing_legend_start, INT1 2026-11-19): keep_from_start beside it, with fmt_block/publish_block taking the element to keep. Only lines past the ordinary 30-line window qualify, so a block that already shows its element renders byte-identically and no capture that regenerated before moves. chooseInWindow and its eight named steps are UNCHANGED -- with the cut keeping the scoped page its first pass finds a window on its own: "--in= case window derived as HEAD~6 (margin HEAD~5 holds)". inCaseUnshown takes the same keepFrom the published block is formatted with, so what the probe measures stays what a reader sees. Disclosed on both sides, because the elided part moved from the end to the middle: the preamble's "How to read the blocks" paragraph says one block keeps a trailing element and that the elided lines there are the middle of the preceding global block, and the case's caption says so when the kept tail was needed -- measured per run by re-rendering the same window with the plain cut. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… not only when it starts with one Found once the generator could write at all (the previous commit), by running the capture's own coverage check against the regenerated document: arm (G) went red with "caption claims [adaptive but the block has no such clause" on --for="tree-sitter parse of a source file" --adaptive. Not a behaviour difference between the binaries -- both emit the clause, verified by crossing tree x binary on the same tree (merged 37da0ae and main ee8d1a5 each print it once). The clause had MOVED: with main's binary it sits on display line 2, inside the head comments the preamble promises in full; with the merged binary the answer-first --for document carries its legend in the TAIL. And the re-wrap could not see that tail. The seam test was startswith("<"), so the raw line that CLOSES such an answer -- `}]]></b></fold><tail total=…><t …></tail><!-- legend --></ctx>`, beginning with the CDATA body's own `}` -- was left whole: one 300-byte-truncated display line holding the entire tail, legend included, which the 30-line window then dropped and trailing_legend_start could not keep because it never saw it as lines. MEASURED on this repo: 91 display lines against 164 once the seam test looks inside the line, where the final six ARE the trailing legend and one of them carries [adaptive: kept N of N]. So the test now accepts any long line that CONTAINS a seam. The JSON branch is tested FIRST so an object line keeps its own seams, and a line under the 300-byte cut is untouched either way. The direction is strictly more content published: the old path showed 300 bytes of that line and nothing else of it, the new one re-wraps it and the ordinary window and marker apply. This is a PRE-EXISTING generator defect, not a regression of the previous commit: a keepFrom=None block takes min( trailing_legend_start, len(lines) ), which is what it took before. The refusal was hiding it -- nothing had regenerated this capture against an answer-first binary. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…the merged build The merge of main took main's side of the dated capture and marked it REGEN AFTER BUILD (IN3), so neither side's capture fitted the merged binary: main's showed --lsp-langmap and missed the stack's three new flags, the stack's was the reverse. The generator refused to re-record on this tip until the two previous commits; this is the re-record. 270 commands, exit 0, recorded on a CLEAN tree at bc21766 (the first run was discarded because an uncommitted generator edit made it read at="…+dirty"). The --in= case's window was DERIVED, not pinned: "--in= case window derived as HEAD~8 (margin HEAD~7 holds)". test/showcase_coverage_check.py against this capture and this binary: PASS (D) coverage: 192/192 flags covered (heading or Not-run) <- was 3/192 uncovered PASS (E) caption-vs-error PASS (F) contrast-pair: 24 pairs PASS (G) header-clause: 2 claim(s) checked, all fulfilled <- the seam fix docs/COMMANDS.md is regenerated from it in the same commit, because docscommandscheck arm (G) requires the shipped document to be reproducible from the committed capture plus the binary: 195 flags in 7 sections, 176 samples, and --check reports the documented set equals the binary set. The "Not run (and why)" list is UNCHANGED -- no flag was moved into it to make an arm pass. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…unt is disclosed as a floor The ratchet counted src/ingest_crawl.h:124 -- the `//` sentence explaining why Metal rides on Lang::Cpp -- as the 326th shared-file `== Lang::` predicate against a ceiling of 325. The line is identical on main and on this branch and no side added it; it became VISIBLE because the grep verb now serves, with their floor disclosed, the hits it used to withhold in a parse_degraded file. In such a file no span tier is known, so --grep-in=code cannot exclude a comment there: measured, that flag's listing is byte-identical to the default's and both count 326, so passing it on the ratchet command would have been a no-op. The fix therefore belongs in the counter. count_predicates now skips a hit whose line begins `//` or `/*`, alongside the `case Lang::` rejection it already had. `*` is deliberately not a comment opener: src/ holds real code lines that start with one and compare (`*okOut = rc == 0;`, src/diffhunks.h:135), so rejecting a leading `*` would one day drop a genuine `*lang == Lang::Cpp`. DIRECTION, stated because this is a bar-adjacent edit on a shared gate: it DOES lower the count from 326 to 325 and make the gate pass. It is legitimate only because it removes a FALSE finding and makes the counter measure the subject the ratchet file names -- a comment is no more a predicate than a `case` label, which the counter already excluded. The ceiling is NOT raised: the ceiling row and the six perlang rows are byte-identical to 2d03f0b by sha256, and the governed quantity (325) equals it. The comment is NOT reworded, because that would pin the output and the next prose predicate re-breaks it. Proved against generated listings, not edited source: the same saved --regex listing counts 326 with the previous counter and 325 with this one, and an 11-arm fixture keeps every real-code shape counted (==, !=, a trailing // on a code line, a tab-indented predicate, a *-dereference predicate) while rejecting the three comment shapes, with the per-language exclusion untouched. Ruling 108, the same counter's second defect, is DISCLOSED here rather than fixed: the hit regex requires `in=`, which the tool omits for a hit with no enclosing symbol, so such a row is dropped silently -- MEASURED, 12 real `l == Lang::X` lines in src/langfacts.h:544-555 are counted by nobody and the printed 325 stands against 337 served rows (IN5's source-level 339 says at least one further line is not served at all, so 12 is itself a floor). The gate's own PASS/FAIL line, ratchet.tsv's header and langidentitycheck.sh's contract paragraph now all call the figure a FLOOR and say the ceiling's quantity is wrong in the SAFE direction -- too low, so the bar is stricter than it reads. The ceiling is deliberately not re-based: making `in=` optional lifts the figure to 337 at once, and that file's rule is "may fall and may never rise". Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…fixbudgetcheck's H2 cell, with a REVISIT each Two owner-ruled acks (104, 105) for integration/070-1120. Neither gate has an ack or waiver mechanism -- the repo's only ack ledger is .ripwire_quality_acks, which keys quality-delta findings, and no gate reads a known-failure list -- so each ack is RECORDED, in the header of the gate a reader opens when the red appears, and both gates STAY RED. Nothing was relaxed, no arm or grid point was removed, and .ripwire_quality_acks is untouched. loopconservationcheck (ruling 104). Arm (A): `2550bfdcd752 (parent e6671f2): 1 entry vanished from the loop with no '# retired:' tombstone: phantomdefcheck`. THE MERGE is "Merge branch 'lane/v1-verbdef' into lane/rt1-route", on the stack's shared history -- in integration/070-candidate, lane/cb1-code-clean and lane/bx2-fix2-ceiling alike, absent from main, which is why main passes and this branch does not. The gate is correct about history; the tree it guards is right today (the retirement is declared at test/regression.sh:290 and test/phantomdefcheck.sh exists on neither tree); and it cannot be otherwise without rewriting that merge out of the history (forbidden) or changing the conservation rule itself to let a later commit's tombstone discharge an earlier vanish -- which needs its own red-first arm and is not an integration branch's edit. Every mutation control (B..I) is green on the same run, so the walker is live and this red carries information. REVISIT when that rule is designed, or when this history is no longer an ancestor of what ships. w3fixbudgetcheck H2 (ruling 105). ONE of 24 points, and the cell is named so nobody re-derives it: lens --pack-task, task length `mid`, budget 3000 -- the top budget step, judged against the 1600 step below it -- `the 9 row(s) served do not extend the 4 served at the budget below -- RP is not a growing rank prefix`. The other 23 points hold, non-vacuity is green at 19 of 24, and main passes the whole arm, so the behaviour travels with the --pack-task D1 rebuild on the cb1/r1b chain and not with the merge. RP is a property of --pack-task's own per-section share assembly, which this gate itself calls a SOFT target; making the 3000-budget page a rank prefix of the 1600-budget one is a product change owned by that lane. REVISIT with the --pack-task section-share owner as part of 0.7.0's budget work. Both failures were reproduced locally on this tip before the acks were written, and both files took pure INSERTIONS: a line-level sha256 comparison against 2d03f0b reports removed=0 for each. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The single 0.7.0 integration branch: the answer-first stack reconciled with
main.Two forked lines of 0.7.0 work plus
main, merged into one branch. No rebase, no squash, nocherry-pick — the merges carry the files in rather than replaying patches against a tree that lacks
them.
What this brings
144non-merge commits overmain(175including merges). Of those, the part this PR had toreconcile is 42 commits:
integration/070-candidateandlane/r1b-priorityshare a merge base2f5561c26that already sits 54 commits pastmain, and diverge from it by only 23 and 19. The rest isshared stack history that was already reconciled below that base —
lane/af1-answerfirst(AF2) among it,which is an ancestor of the base and so needed no landing step of its own.
--cherry-markfindszero duplicate patches between the two lines.
By theme:
--for— the paged answer (--page-bytes=), the key-body block(
src/forkeybody.h), the signature index and the page layout (src/forpages.h,src/forsigx.h),and the page budget's explicit priority order.
--pack-taskD1 rebuild — the--forhead, the file tail, one key body, one token conversion;the tail's declared share is a floor; a truncated served body is disclosed as a cut.
kKeyBodyMax = 5(was 2) under a new monotone120-line cap,
kCappedBodyLines.--forrouting (RT3) — call-path and quoted-literal questions are vetoed; a windowed routedanswer continues itself.
--impactfirst window (IM1) — production rows and evidence importers lead it.--callers/--usessee-through.defGroupsRenumberMatchescollects a dropped group's member so aprinted table still lists it.
SubagentStartprimer hook, the--legend-seen=flag,--no-verb-route, one aggregateci-gatejob, and gate hygiene.main: the--lspclient-declared-language path and its--lsp-langmap=alias map, and thetrain-1117d1 extraction work.
How the merge was done
Union first (
integration/070-candidate×lane/r1b-priority, zero human conflicts), thenmainby shaee8d1a50ea9be710c9b7b6a875b1afa7073d074c— by sha, not by name, so a concurrentfetch cannot move the target. Both merges went through the repo's own
scripts/merge_main.sh+scripts/merge_resolve.py, which set-mergetest/regression.sh's absorb list, settle the gatecountsites, mark the binary-pinned files REGEN AFTER BUILD, and refuse a
kParserVerthat does not clearboth sides' bumps.
mainleft 13 files / 17 hunks for a person. The ones worth a reviewer's eye:src/defattrib.h— the only genuine two-sided refactor collision.mainsplitdefGroupsDropSilentintodefGroupsSilent+defGroupsDrop+ a thin wrapper; the stack kept themonolith and added
defGroupsRenumberMatchesabove it, so the naive merge left two definitions ofone function in one header, which cannot compile. Resolved to
main's structure with the stack'shelper kept, because the stack's only semantic change to the monolith was replacing its
ofMatchloop with that helper — and that loop lives, under
main's split, insidedefGroupsDrop. Main'ssplit is therefore a strict superset of the shape the stack's change needs. In the second hunk
mainchanged one word of a comment where the stack changed the behaviour, so the behaviour is kept
(
test/verbfixcheck.sharm F7 pins it).kParserVer/kIngestParserVerMirror—main331, the stack 317, base 313. Both moved to332 (
max+1), rich file tag 333, so neither side's cache can be read by the other's binary. Bothprovenance lists are kept as labelled history.
test/qschemetrip.hashis re-pinned from the mergedbuild (
0a4ba8ec92…): neither side's value could be right once the parser version moves.main27/226 (--lsp-langmap=), the stack 28/228. Neitherside taken: 28 / 229, with both provenance lists merged. The three prose "long flags" counts went
to 199, which is not an arithmetic guess — it is
--help=allcounted the gate's own way on thebuilt binary.
skills/ripwire-mcp/SKILL.md— the stack split this skill into 7 files whilemainedited themonolith. The split is kept and
main's two LSP-langmap commits (9b8c57052,b47088b6e) wereported into
transports.md, the split file that owns--lsp: thetextDocument.languageIdpath, the config search order, the
[overrule] suffixesrule, and the explicit-@langexception.test/legendcoveragecheck.shandsrc/mcpverbs.h— independent additions on both sides; bothkept. The gate asserts no file-wide arm total, so the concatenation runs strictly more arms.
test/compactlegendfix/pre_l1_full/callers.xml— a hand-re-anchored pin, not a generated file.Both sides' edits were characterised mechanically and both applied; the result differs from
main'sside by exactly the stack's two insertions and nothing else.
Three gates exist on
mainand on neither line, and a merge taking the stack's side oftest/regression.shwould have dropped them silently.fnvbasischeck,langordercheckandqdrangecheckare all present and registered (two in the absorb list, one invoked by name). Theregistry check proves the stronger statement: all 898 top-level
test/*.shfiles are registered exactlyonce, no duplicate and no orphan — so no gate from either side was lost.
Measured gains (lane R1b, on lane BL1's units and instrument)
Body rung: 9 → 20 of 89 cells complete, 14 cells up, ZERO down, closing 11 of the 21-cell
composition gap BL1 measured. R1b's control arm reproduces BL1's baseline body figure exactly
(9 of 89), which is the strongest instrument agreement available between the two lanes.
Bytes: page 1 never grows by more than +679 B in any cell, 0 of 89 cells exceed the pre-registered
+1,500 B per-cell cap, no cell's whole answer grows at all, and the population sheds 481 B on
page 1 and 2,145 B overall.
Guard
B-FILE-GOLDat zero tolerance: PASS — 122 → 122 gold FILE rows, 0 cells up and 0 cellsdown, and not one of the 11 repos moves in either direction (Newtonsoft.Json, bash-it, bat, black,
eslint, gson, jq, libzmq, restic, sinatra, zod).
That pass must be read with R1b's own caveat, which is the most important one in its readout. The
loss channel is wide open: 33 of 89 cells do lose page-1 tail rows, 773 rows in total. The gold
simply does not live in those rows — this population carries a median of 1 gold FILE row per task. On
the orient-shaped cells that originally failed this guard, 8 of 9 losses persist. So the honest
statement is "the file rung is unharmed on localisation-shaped questions, and the orient-shaped class
is unmeasured here", not "the file rung is safe".
Function rung: +2 to +4 cells under every one of seven scoring definitions, with zero cells falling
under any of them. The level is deliberately not quoted here: R1b re-scored the same answers seven
ways and none reproduces lane BL1's 62, so BL1's grader remains the instrument of record for the
level and the two must not be quoted side by side. Direction and the zero-loss property do not depend
on the definition; the level does.
Three honest caveats
1.
forkeybodycheckis RED, deliberately, and it is a real contract defect — not a gate to arguewith. Owner ruling 87 names the fix:
insertKeyBodiesshould stop passing theservedargument tokeyBodies, so a body already on page 1 is not suppressed. The call is atsrc/forkeybody.h:253(:239is the declaration — an earlier mis-citation, corrected here). It is a one-argument, one-line change
that is unimplemented, identically, on both merged lines — nobody has written it. Owner ruling 95
settles that it lands as a FOLLOW-UP PR with its gate written first, and the reason is not scheduling:
87's gate has to be written red-first so that its green carries information, and that cannot happen
inside a 42-commit merge whose suite is green for other reasons. So the red here is carried knowingly,
not tolerated. Measured cost of leaving it: +2,145 B in one cell.
2. The function rung's LEVEL is not quoted beside lane BL1's 62. See above — seven definitions, none
reproducing it. Only the delta statement is quotable.
3. The programme's headline completeness pair was measured on a DIFFERENT configuration than the one
this PR ships. The frozen candidate binary (
built_from=79052376e) carries the fork-base answer-firstcode:
kKeyBodyMax = 2and no line cap. This branch lands CB1's raise to 5 plus the 120-linemonotone cap. Every figure quoted from that binary therefore describes code this PR does not ship. The
two agree on the body rung at 9 of 89 — and that agreement is expected rather than evidence of
equivalence, since this population's median task needs one body, so a cap of 2 is not binding for most
of it. Owner ruling 94 settles the consequence: the headline pair WILL be re-measured on this
branch's binary before 0.7.0 ships. So this is not an open question but a ruled, pending step, and
until it completes the published pair describes a configuration this PR does not ship. The
re-measurement is a separate lane's work and is deliberately not attempted here — it needs this
binary first.
Verification
Built with
cmake --build build --clean-first -j4on a plain dev configure (never-DCMAKE_BUILD_TYPE=Release, which would compileDEGRADED_PATH_ALERTout).--clean-firstis notoptional here: this merge moves
kIngestParserVerMirror, which is the repo's own worked example of astale object surviving repeated successful rebuilds. Exit 0, zero warnings, zero errors, and
--versionreportsbuilt_from=0ebd18b66, equal to the branch tip.Full gate suite:
898 gates is exactly the registry total, and
tree_writes=0is the runner's own proof that nothingedited the tree while it ran.
Eight failed on the first pass. Three were a procedural error of mine, not the merge: I built at the
main merge and two generated-file commits landed on top, so the binary's
built_fromtrailed HEAD whileno
src/file differed.versioncheck,versionlabelcheckandmcpreloadasancheckwere all correct tofire, and all three pass after the clean rebuild at the tip.
The five real failures, each attributed against an unmodified
ee8d1a50ebuild. None is main's.forkeybodycheck(K3)langidentitycheckloopconservationcheckshowcasecapturecheck(D)w3fixbudgetcheck(H2)--pack-taskRP monotonicity, 1 of 24 pointsThree of these want a maintainer's eye, so the detail is here rather than buried:
langidentitycheckfails on the binary, not on any source change. The shared-fileLang::ratchet is counted through the tool (
runner.py --ratchet-onlyruns the binary's own--regexoversrc/), so I crossed tree × binary in all four cells: main's tree with our binary gives the same326/FAIL, and our tree with main's binary gives 325/PASS. A source-level count of distinct
== Lang::/!= Lang::lines is identical between the two trees, file for file. All three mergeparents carry
ceiling 325, so no side raised it. I did not raise the ceiling — the ratchet filesays it may never rise, and there is no new predicate to move, so the gate's own suggested remedy does
not fit this cause. This is a gate whose subject is the source but whose instrument is the binary, and
it needs a decision rather than a lane's edit.
loopconservationchecknames a historical stack merge (2550bfdcd, absent from main) in whichphantomdefcheckleft the loop before the commit that declared its tombstone. The tree as it stands iscorrect and self-consistent —
test/regression.sh:290carries the# retired:tombstone and the gatefile is absent from both trees — so this is bookkeeping in history, not a missing retirement, and it
cannot be corrected without rewriting history.
showcasecapturecheck(D) reports3/192 flags uncovered: --legend-seen, --no-verb-route, --page-bytes. Both sides carried the same dated capture and it conflicted; the tool took main's side.Neither side's capture is right for the merged binary — main's covers
--lsp-langmapand misses thestack's three; the stack's covers all three and misses
--lsp-langmap. The dictv pair was alreadyre-recorded to the merged binary's. I ran the prescribed generator,
test/showcase_capture.py, and itrefused to write: its own window guard finds no
--since=HEAD~Nin [2,14] that demonstrates whatthe
--in=srccase promises on this merge-heavy tip. That is the generator behaving correctly, so thecapture cannot be re-recorded on this branch as it stands. I did not force it and did not hand-edit the
capture's "Not run" list to silence the arm — the flags are runnable, so listing them as not-run would
be a false record.
Other verification:
xmllint --nooutclean; zero newlines outside CDATA (G4).--quality-delta:rc=2,regressions=111,minor=75,gating=2, withbaseline="ref-pair"against
ee8d1a50e. The ref-pair form is deliberate: on a committed tree the bare--quality-deltacompares the tree against its own HEAD (head_basis="identity") and reportsregressions=0whatever the branch holds, so that green is free and is not quoted here. Both gatingrows are inherited, not introduced:
packBodiesis byte-identical tolane/r1b-priority's andtest/testrowpaths.pyis unchanged against it, and the aggregate 111/75 is 144 commits from manylanes, each already accepted on its own.
--clean-firstwith-DRIPWIRE_ASAN=ON, exit 0, zero warnings, and six runsclean under
-fno-sanitize-recover=all. The runs were chosen to actually enter the changed code: thedefault map carries no answer-first marker at all, while five
--for/--pack-taskruns at--page-bytes=12000/8000/4000/2000 and a--pack-taskrun all showpage="1",deferred=,capped="1"and the<bodies>block. What ASan does not reach, stated rather than glossed: under asanitizer the CMake condition links neither allocator TU and leaves
RIPWIRE_MEM_CEILING_HOOKundefined, so ASan structurally cannot exercise the memory ceiling this branch carries — that path's
evidence is
memceilingcheckin the plain build's suite, which passed, not the sanitizer run.std::length_error, nobad_alloc, no sanitizer regiondivisible by an old
sizeof(Symbol).Not in this PR
snapshot, not a per-branch value —
merge_main.shsays so in its own header. On this branch thepublished gate count still reads 876 (dated 2026-11-19) where the registry now names 883, and
docs/LIMITS.mdstill reads 447 caps against a live 452. Both were verified not broughtcurrent by this merge, and both generators'
--checkpasses, so the tree is self-consistent as itstands. Refreshing them (
python3 docs/counts_snapshot.py refresh) is a separate pass after thislands — a regenerated published count that later proved wrong should not be attributable to this
merge. What is in this PR is the merge's own gate-count regeneration, which
manifestcheckrequires; that was regenerated, never hand-edited.
🤖 Generated with Claude Code