RAP reproduction: refresh stale scope/results and add an n=98 dominance regression

[arniber21]

What to build

Refresh the RAP reproduction harness, methodology, and regression gates after the #300 and #304 compiler changes, then add a deterministic n=98 dominance check that detects the drift found in #485.

A current release reproduction at commit 268ca3b4ea30, repeated twice, produced:

Fixture	Placer	Steps	Move time	Search
n=42	agnostic	23	3049 us	n/a
n=42	aware	18	2999 us	4 completed / 0 fallback
n=98	agnostic	40	6262 us	n/a
n=98	aware	48	8896 us	0 completed / 4 fallback

The checked-in documentation still records n=98 as 48/48. The sweep also says #304 OpenQASM ingestion is unimplemented even though #304 is closed and ising_n42.qasm reproduces the Quon fixture exactly. Its CSV banner says the enforced n=42 metric test is not invoked by CI, while just ci-rust now runs it explicitly.

This ticket should correct provenance, expand the harness to the implemented ingestion scope where fixtures are available, and ensure future compiler improvements cannot silently make aware fallback worse than agnostic.

Acceptance criteria

• Methodology and CSV banner report the current n=42/n=98 values and identify the compiler commit/tool version that produced them.
• All claims about NA: OpenQASM ingestion for the neutral-atom pipeline (NA-scoped slice of #197) #304/QASM ingestion and just ci-rust wiring match the executable repository state.
• The n=42 OpenQASM fixture is included in an equivalence check against the Quon fixture for structure and schedule metrics.
• The harness records agnostic mechanism per layer or an equally precise mixed-mechanism summary.
• n=98 has a regression asserting that routing-aware fallback is no worse than the current agnostic dispatcher under the same target/objective.
• Published RAP values remain separate comparison columns; current Quon values are never relabeled as paper reproduction numbers.
• Runtime, memory, search-budget, speed-model, and transfer-time caveats are consistent across the script, methodology, Justfile, and validation docs.
• The sweep emits sufficient provenance for repeatability: git commit/dirty flag, compiler version, target ID/hash, fixture hash, configuration, and deterministic seed if any.
• Local/nightly versus required-CI rows are stated once and checked by a test where practical.

Out of scope

• Changing target constants or fixtures to make results numerically closer to RAP.
• Adding paper benchmark inputs whose licensing/provenance has not been reviewed.
• Fixing the fallback algorithm itself, tracked by NA/zoned: make routing-aware fallback dominate the current agnostic placer #485.

Blocked by

• NA/zoned: make routing-aware fallback dominate the current agnostic placer #485 — fix routing-aware fallback drift before locking the new n=98 dominance baseline

Related

• Reproduce a published TUM benchmark number as a checked-in regression #111 — n=42 regression anchor
• NA/zoned: implement RAP A* heuristic (Eqs. (3)–(5)) so routing-aware placement completes, then enforce Table I #297 — aware heuristic
• NA/zoned: vertex-matching placement + reuse analysis for the routing-agnostic baseline (Lin 2025 parity) #300 — matching-based agnostic placer
• NA: OpenQASM ingestion for the neutral-atom pipeline (NA-scoped slice of #197) #304 — OpenQASM ingestion
• NA benchmarks: full RAP Table I sweep + qmap-comparable CSV harness #306 — sweep harness