M5: SI/PI/thermal/DFM surrogate hooks with authoritative signoff

[seunghyukchoe]

Surrogate models may rank candidates but never sign off: any SI/PI/thermal/DFM claim surfaced to a caller must come from an authoritative tool run (or be declared a non-claim), keeping the 'AI proposes, deterministic code disposes' invariant. Design the hook contract first: where surrogates plug into scoring, what evidence binds a signoff, and what the refusal looks like when no authoritative backend is configured. Roadmap: docs/roadmap.md § M5.

[seunghyukchoe]

Parked behind #99, staying in M5 (now M5 — Verification and physics). Not closed: the contract
this issue asks for is still the right thing to design, and designing it first is still the right
order.

Parked because the half of this issue that carries the safety property is the signoff, not the
surrogate — "any SI/PI/thermal/DFM claim surfaced to a caller must come from an authoritative tool
run (or be declared a non-claim)" — and that is the same contract #99 has to define first: what it
means for CopperMCP to accept or refuse someone else's proposal, with typed diagnostics and bound
evidence, and what the refusal looks like when no authoritative backend is configured.

Building a surrogate hook before that contract exists means designing the ranking side against a
disposal side that is still a stub. Verified: validate_candidate
(src/copper_mcp/tools.py:408) normalizes a manifest and returns {"valid": True, …} — no
geometry, no verdict — so there is currently no shape for a signoff to bind to.

Once #99's first slice lands, the evidence-binding vocabulary it establishes is what this issue's
hook contract should reuse rather than reinvent.

Recorded in the post-0.8.0 audit and improvement plan as item P5.6.

[seunghyukchoe]

DFM executor slice merged in #213 (merge commit 2036b87). Production-core execute_dfm_signoff now invokes only the fixed KiCad DRC path, requires 2-8 exact canonical observations to agree, refuses suppression/disagreement/failure/cancellation/cooperative deadline stops, and returns redacted candidate/revision-bound results. The seam remains absent from MCP/CLI/apply.

#91 stays open: SI, PI, and thermal remain unregistered non-claims; surrogate hooks and any reviewed public exposure are not implemented. The module-private Python capability is explicitly a cooperative supported-path guard, not a hostile same-process sandbox. Local make check: 3,284 passed, 2 skipped; all 8 PR checks green. PR review found and remediation covered workspace-boundary path failures.

[seunghyukchoe]

Predeclared measurement: surrogate-top-pick DFM signoff agreement (instrument PR to follow; full run is a separate evidence commit).

Question: on a fixed committed corpus, does the private surrogate's top-ranked variant sign off (DFM, repeated agreement) more often than its bottom-ranked variant?

Corpus (committed bytes only, no imports): benchmarks/audio/fixtures/ne5532-stereo-summing-routing-v1.kicad_pcb, negotiated-crossing-v1.kicad_pcb, rc-low-pass-routing-v1.kicad_pcb, tests/fixtures/route-candidate/two-pad.kicad_pcb, partial-route.kicad_pcb, tree-star.kicad_pcb.

Rules fixed in advance: every net with >=2 selected-layer pads is attempted; K=4 variants per net at seeds (101, 202, 303, 404) via the reference A* core; nets where all variants refuse are excluded and counted by refusal reason; nets with fewer than 2 routable variants are excluded and counted. Rank each net's routable variants with rank_surrogate_candidates (DFM domain, advisory only). Sign off exactly top-1 and bottom-1 per qualifying net with execute_dfm_signoff repetitions=2 (the comparable minimum), DFM domain only.

Metric: top_signoff_rate minus bottom_signoff_rate plus the full per-net table, descriptive only. Stop rule: report only -- no tuning, threshold, or follow-up slice follows in the measurement commit regardless of outcome, and no SI/PI/thermal, MCP/CLI, live, apply, or fabrication claim attaches. Pilot (unpublished): negotiated-crossing full pipeline to validate cost; full run binds the final runner and publishes a self-digested aggregate artifact with its own B-row.

[seunghyukchoe]

Clarification to the predeclare above (before any run): variant outcomes are ternary -- routed (candidate), already_connected (nothing to rank or sign off), or refused with the core's reason. Only routed variants enter ranking and signoff; already_connected variants are tallied separately under their own count. The 'fewer than 2 routable variants' exclusion counts routed variants only. Profile for all variants: clearance 250000nm, track 250000nm, via diameter 800000nm, via drill 400000nm on layer F.Cu under default AStarSettings; fixed in the runner, documented in its module docstring.

[seunghyukchoe]

Amendment to the predeclare (pilot evidence, before any full run): the pilot reached signoff execution on both fixture nets and every pick refused suppressed_evidence -- the fixture board's plain KiCad DRC skips 5 checks in a bare workspace (verified: passed False, ignored 5), and ADR-0119 refuses skipped-check runs even when they pass. A full run without a reachability gate would therefore report a trivial 0.0 differential that measures workspace context, not ranking signal. Amended rule, fixed before the full run: each board first replays its plain board DRC; boards with ignored_check_count or exclusion_count nonzero are tallied as board_context_ineligible with routing and ranking still executed offline, and signoff runs only on eligible boards. Nets on ineligible boards are excluded from signoff denominators under their own counter, never folded into the fewer-than-two bucket. The pilot already validates the amended pipeline shape.