Baton: the #332 declared-signal principle applies to the Coherence Bridge — codescan's role=language heuristic is the same anti-pattern as title-scan

[montfort]

Context. Baton has two axes that the recent #332 turn touched very differently, and they share one underlying principle worth recording explicitly — so that when the next piece of work on the Coherence Bridge lands (notably #321, triggered at N=2), it is done under the lesson of the turn, not by re-deriving it.

This is a design-coherence note, not new functionality. It does not change scope already owned by #321 (the codescan generalization plan + its N=2 trigger). It records the why that should govern that work.

The two axes (for the record)

• Cost-Aware Router (Phase 2, CHARTER-03) — "what kind of work is this unit, to route it to which model tier?". Reads governance metadata, never source code.
• Coherence Bridge (Phase 1, CHARTER-01) — "does the code that exists match the contract the spec/decision declared?" (the Cross-spec decision propagation: a post-MVP backend decision silently diverged a later consumer spec's contract — the system view should link them #304 case). Reads real code via experiment-baton/src/codescan.rs.

The shared anti-pattern: reconstructed-heuristic-calibrated-to-one-stack

The #332 turn discontinued the title-scan on the router axis. The root reason (decision #332, AILOG-2026-06-27-001) was not "keyword tuning is hard" — it was structural: inferring a unit's work type by substring over an unconstrained, schema-less title is calibrated to one adopter's titling habits, an unbounded per-adopter treadmill (the #321 "calibrated to one stack" anti-pattern). The fix was to prefer a declared signal at authoring (work_verb) over a posteriori reconstruction.

That is exactly the load-bearing smell #321 already identified on the Bridge axis: codescan hardcodes Go = producer, conflating language with architectural role. Role is not a property of language — it only looks that way because Sentinel's stack happens to be Go⇄TS. It is the same error as title-scan, in the other axis: a heuristic calibrated to the single dogfood oracle.

Both are the reverse of the same thesis (#328): prefer declared / explicitly-confronted structured signal over heuristic reconstruction.

• The router now classifies declared work (work_verb) instead of guessing it from the title.
• The Bridge already keys contracts by the normalized endpoint (scan::normalize_endpoint) — a declared, language-neutral anchor — and links intention to governance by explicit IDs, not textual similarity. That core is right. What is not aligned with the thesis is the stack-specific extraction around it (file-by-extension, Go = producer).

What this means for the Bridge (direction, not a redesign)

1. Baton: language-agnostic contract-shape extraction (codescan is Go⇄TS-specific; decouple role from language) #321's generalization should be governed by this principle. When N=2 triggers it, infer producer vs consumer from signal (a serialization-tagged struct reachable from a route registration = producer; a type consumed at an HTTP call site = consumer), never from language. Follow the precedent the turn endorses and that Baton: language-agnostic contract-shape extraction (codescan is Go⇄TS-specific; decouple role from language) #321 already names: the config-driven, language-agnostic scanner pattern Loom shipped (architecture: in config.yml, architecture generate: language/structure-agnostic scanner — configurable extensions + container dirs (Java/Maven collapses to one 'main' box) #279) — i.e. let the adopter declare producer/consumer source globs + languages, rather than baking a stack assumption.
2. Baton: producer-side route keying for handler-framework Go (huma route-registration ↔ output struct) #319 is consistent with the thesis and remains valid — it fixes structural reading (huma route-registration → output struct), not a keyword table. No change.
3. Open question the turn raises (defer with Baton: language-agnostic contract-shape extraction (codescan is Go⇄TS-specific; decouple role from language) #321, do not act now): should the Bridge prefer a declared source of contract shape (OpenAPI / contract frontmatter) over heuristic code extraction, by the same cost/certainty argument as Decisión: descontinuar title-scan; work_verb declarado como única señal de clasificación #332? Baton: per-type → endpoint contract keying for generated type files (unblocks C2/C3) #313 chose "call-sites > parsing OpenAPI"; that choice is worth re-examining in the turn's light — but only once a real second stack is on the table.

Bridge ↔ routing (clarifying the relation, since it motivated this note)

By design (concept §10.6): "the router consults coherence before routing" — coherence first, routing after (§3: a cheaper router over forgotten intent just automates drift faster). But that consult is design, not code today: the two subsystems are separate; the router does not consume Bridge findings. The concrete integration (a unit touching a drifted contract = risk signal → route up) was among the router's deferred heavy signals. The #332 turn established that the confidence lever is declared signal, not wiring those heavy signals — so Bridge→router integration stays a secondary/future cross-check, not the primary routing mechanism. Recorded here so the relation is not misread as an existing dependency.

What this issue does / does not do

• Does: record the shared principle so Baton: language-agnostic contract-shape extraction (codescan is Go⇄TS-specific; decouple role from language) #321 (and any future codescan work) is executed under the turn's lesson; make the two-axes relation explicit.
• Does not: re-scope Baton: language-agnostic contract-shape extraction (codescan is Go⇄TS-specific; decouple role from language) #321 (its generalization plan + N=2 trigger stand), change Baton: producer-side route keying for handler-framework Go (huma route-registration ↔ output struct) #319, or wire Bridge findings into the router now.

Relation

#332 (the turn — title-scan axis) · #321 (codescan role=language — Bridge axis, the generalization) · #328 (structured-signal-is-the-lever thesis) · #333 (title-scan removal) · #304 (the umbrella the Bridge serves) · #319 (producer-side keying, consistent).