fix: inject resolved rule contents into the run-stage directive (#495)

[wowzoo]

Fixes #495.

buildRunStageDirective ships rules_in_context as paths only, and nothing in
the engine forces the conductor to read them. Per-stage steering could therefore
be skipped silently — the paths listed, the files present with substantive
content, the artifacts written without them.

The engine already has the pattern this needs, twelve lines below the defect:
conductor_persona is delivered by reading aidlc-common/conductor.md at
directive-build time and baking in its contents, so no skill references it by
path. This applies the same treatment to the rule layers.

The change

• aidlc-directive.ts — new RuleContent type; optional
  rules_content?: RuleContent[] on run-stage, registered in
  RUN_STAGE_FIELDS, validated by checkOptionalRulesContent (shape discipline
  mirrors checkOptionalConsumesAbsent).
• aidlc-orchestrate.ts — readRuleContent reads each resolved path
  best-effort; absent, unreadable, or placeholder-only yields no entry. Gated on
  codekbCtx so the ctx-less emit path stays byte-identical, mirroring
  splitConsumesByPresence's own no-ctx skip.
• ruleTextIsSubstantive drops files whose every line is blank, a heading, a
  > blockquote, or an HTML-comment fence — the exact shape of the shipped
  team.md / project.md seeds, so an untouched install gains no directive
  noise. They appear as soon as they carry a real rule.

rules_in_context is unchanged and remains the authoritative roster;
rules_content is an additive subset over the files that carried content.

Limitation — this reads the default space only. readRuleContent resolves
each entry of node.rules_in_context, and those paths are baked by
memoryDisplayPath() whenever the graph is compiled — pinned to
MEMORY_SPACE = "default", as is the rulesDir() the roster is walked from. Nothing
on this path consults the active-space cursor, so with a cursor on teamB the
injected content is still aidlc/spaces/default/memory/. That pin is the shipped
COMPILE/DISPLAY-family behaviour documented in aidlc-graph.ts and this PR does
not change it — but it is worth stating plainly, because injection makes the
default content deterministic rather than merely listed, and the conductor
demonstrably stops reading paths it has already been given (see the run below).
Making rule resolution cursor-aware is a separate, harness-wide question.

No skill or protocol prose changes — and a live run explains why

I first rewrote the conductor's "read every file in rules_in_context" clause in
all five harness SKILL.md files, then measured whether it was needed. It is
not, so I reverted it and this PR touches no prose.

A live Kiro IDE 1.0.212 run, instrumented with a temporary PreToolUse probe on
read_file|read_files (the framework's own hooks cannot see reads — their
matchers are write/shell only), produced this in a single read_files call at
intent-capture:

rule file	in rules_content	conductor read it
org.md (5439 chars)	yes	no
phases/ideation.md (1158)	yes	no
team.md (placeholder)	no	yes
project.md (placeholder)	no	yes

The read boundary matched rules_content membership exactly: the conductor
skipped precisely the two files whose contents it had already received, and read
the two it had not — in the same call as the stage file and stage protocol, so
this was not laziness. Injection adds no duplicate read, and the existing prose
needs no change.

Tests

New t248-directive-rules-content.test.ts, CLI-boundary (the builder has no
exports), 7 tests: content present and verbatim, subset-of-roster, placeholder
exclusion, a substantive team.md appearing, an emptied org.md dropping out,
the phase rule tracking the stage's phase, and the no-memory-tree case leaving
the field absent with routing unaffected. Registered in the coverage registry's
none->cli spawner list (the honesty ratchet requires the explicit edit).

Verification

• bun run check — green (parity 5 harnesses, typecheck, lint 540 files)
• bun tests/run-tests.ts --smoke --unit — 178 files, 0 failed, 4395 assertions
• bun scripts/ci-changelog-guard.ts <base> — OK, 137 preserved, 1 new
• Fixture emit: org.md + phases/ideation.md injected, both placeholder seeds
  dropped; adding a rule to team.md makes it appear immediately

The integration tier's t66 and t89 also fail on base 257b43a3 (2.5.11 added
the claim-sources sensor without refreshing the designer-export golden fixture
or the sensor-import fixture dir). Both are outside the PR gate's tiers and
unrelated to this change; happy to file that separately.

Version bumped to 2.5.14 (2.5.12 is #615, 2.5.13 is #653) with a matching
CHANGELOG entry and README badge. docs/roadmap.md drops #495 from "Known gaps".

Scope note: the two edited files are shared core — byte-identical in all five
dist/ trees — so the behaviour lands on every harness, not just one. Worth
naming the asymmetry: the harnesses with a native include (Kiro CLI's agent
resources glob, Claude's @-import stub, Codex's AIDLC_RULES_DIR,
opencode's instructions) already carry the memory tree in ambient context, so
for them this is a per-stage determinism guarantee over a roster the glob does
not distinguish rather than newly available content. On Kiro IDE, which has no
such native include, the injected content is the deterministic path. In both
cases the determinism is over the default space, per the limitation above.

[apackeer]

Thanks, Keehyun. The diagnosis and initial implementation here were correct: required rule content must be delivered by deterministic tooling rather than left to discretionary path reads. I ran the focused test from this exact head; all 7 submitted cases passed.

Since this opened, #658 merged that same core direction into current v2 and carried it through the additional failure paths found during review. The shipped implementation resolves the active space, fails closed on missing, unreadable, or invalid-UTF-8 required rules, preserves authored blockquoted policies, splits large bundles into integrity-bound load-steering parts, and enforces delivery at dispatched-worker boundaries. It also absorbs reviewer checklists into their generated agent bodies. The current focused suite passes all 32 cases.

The separate persona and supplemental-knowledge retrieval gap you called out is not lost; it is tracked in #694. Issue #495 itself is now resolved by the merged 2.5.33 implementation, although it remains open because changes merged to v2 do not auto-close issues.

Please do not spend time rebasing or reworking this branch. I recommend closing #654 as incorporated by #658 and manually closing #495 as resolved. Thank you for the live evidence and initial patch that established the direction.