feat(adf): real module-attribution for METRICS constraint failures

[stackbilt-admin]

Summary

Implements real module-attribution for charter adf suggest's loadedButViolated detector, replacing/augmenting its previous session/time-window correlation heuristic with exact evidence for METRICS ceiling failures.

Scope correction first: earlier work in this thread assumed validate/drift/adf-evidence all needed module-attribution. That premise was wrong — validate.ts checks git commit governance trailers and drift.ts checks blessed-stack anti-patterns in source files; neither is module-shaped, so there's nothing to attribute there. Only adf-evidence.ts's METRICS ceiling checks (and serve.ts's getProjectState, left untouched since it already groups its own output by module) actually check per-module ADF constraints. This PR closes that gap only — verified via advisor review before implementation.

• ConstraintResult gains an optional module field; validateConstraints takes an optional third module param.
• BundleResult gains an optional documents map (module path → parsed doc); evaluateEvidence validates per-module when available, attributing constraints back to their module, falling back to the old merged-document (unattributed) path when documents is absent — so hand-built BundleResult fixtures in existing tests are untouched.
• New AdfConstraintEvent telemetry type + recordAdfConstraintEvents(): one event per failing, module-attributed constraint, written by charter adf evidence.
• adf suggest's loadedButViolated now reports exactFailures (real evidence, joined by module + session/time) alongside the existing correlatedFailures (kept as the fallback for non-module-shaped checks and pre-existing telemetry).
• packages/adf and packages/cli bumped 1.8.0 → 1.9.0 (additive minor).

All changes are additive — no public API removed or renamed, no breaking changes to existing BundleResult/ConstraintResult consumers.

Test plan

• pnpm build / pnpm typecheck clean across the workspace
• 494/494 tests pass (packages/adf + packages/cli), including new coverage: per-module attribution + merged-doc fallback (evidence.test.ts), documents population (bundler.test.ts), exact-vs-correlated join logic (adf-suggest.test.ts), and the adf.constraint event write/skip paths (evidence-contracts.test.ts)
• Verified live (not just fixtures): ran a real adf bundle --task ... → adf evidence --auto-measure (breached ceiling) → adf suggest sequence in a scratch repo with CHARTER_SESSION_ID set and confirmed exactFailures: 1 on the real offending module
• This repo's own pre-commit adf evidence gate passed cleanly during commit (dogfooding the exact code path this PR changes)

🤖 Generated with Claude Code

[stackbilt-admin]

Addressed all 5 findings from the code review:

1. exactFailures count inflation (fixed) — was counting joined resolution events instead of distinct constraint failures, so a module resolved 5x in one session with a single real ceiling breach reported "5 confirmed constraint failures." Now attributes exactFailures directly from the number of fail events per module, independent of resolution count. Verified live: 5x adf bundle + 1 real breach → occurrences: 5, exactFailures: 1.
2. bundle.documents truthiness gap (fixed) — bare truthiness check treated documents: {} as "per-module data available," which would have silently produced zero constraints instead of falling back. Added a hasCompleteDocuments guard that confirms every resolved module has a matching document before taking the per-module path.
3. Duplicated aggregation logic (fixed) — extracted computeConstraints() as the shared per-entry validation logic; validateConstraints wraps it with the aggregate fields, evidence.ts's per-module path uses the lighter helper directly instead of discarding a redundant weightSummary/failCount/warnCount per module.
4. Parallel documents/documentsByPath structures in bundler.ts (fixed) — removed the array; mergeDocuments now takes Object.values(documentsByPath).
5. Three separate Maps + Set-union in adf-suggest.ts (fixed) — consolidated into one Map<string, LoadedModuleStats>; added a shared groupBy() helper used by both the new constraint-event grouping and the pre-existing co-occurrence grouping.

Added regression tests for #1 and #2. 497/497 tests pass, full workspace build/typecheck clean.