Spec audit: docs vs code for the CTS compiler surface

Why this exists

The 2026-07 audit found the current-behavior spec denying an implemented,
fixture-pinned CFC ifc.* lowering subsystem. Use this map to distinguish
normative disagreement from descriptive drift and to find the repository's
load-bearing evidence quickly.

The map

• Start at docs/specs/ts-transformer/README.md for the corpus and each
  document's authority. Normative target-language/lowering contracts win over
  code; descriptive current-behavior and schema-mapping specs follow code and
  passing tests.
• Ground implementation claims in the relevant canonical constant or source,
  targeted tests, and—where output shape matters—the
  packages/ts-transformers/test/fixtures/** input/expected corpus. Inspect
  live emission with deno task cf check <file> --show-transformed --no-run.
• Canonical constants the specs cite (spot-check the citations still resolve):
  CFC_TRANSFORMER_STAGES (cf-pipeline.ts),
  COMMONFABRIC_RUNTIME_EXPORT_REGISTRY (+ its runner-factory guard test),
  CrossStageState, getExpressionContainerKind,
  SES_SELF_CONTAINED_CALLBACK_BOUNDARIES, wrapper vocabulary in
  schema-generator/src/typescript/wrapper-names.ts,
  CFC_CANONICAL_ALIAS_NAMES in packages/api/cfc.ts, and the JSONSchema
  type in packages/api/index.ts. Mechanical stage claims are pinned by
  packages/ts-transformers/test/spec-sync.test.ts.
• Cross-package contracts with no single home (historically the blind spot):
  transformer emission ↔ runner sandbox verifier (@commonfabric/utils sandbox
  contract); pattern-coverage ↔ runner line-offset/cache couplings
  (runner/harness/engine.ts); bare WeakMap boundary with schema-generator
  (cross-stage-state.ts header); CFC aliases/policy markers ↔ schema generator ↔
  policy compilation ↔ runner fail-closed IFC validation.

Known drift patterns (seeds, not the boundary)

• Landing-gap: a feature ships, spec sections that deny it stay (grep the spec
  for "not implemented", "no …yet", "draft" and test each such claim).
• Enumeration drift: prose lists (diagnostic ids, registry entries, stages,
  fixture suites) vs their constants — including the inverse direction (ids in
  code the spec never mentions).
• Rename shadows: file/symbol references in older docs after renames; git touch
  dates lie (mechanical rename PRs refresh mtimes without verifying).
• PR-scoped artifacts aging in place; "Effective date"/status headers nobody
  updates.
• Author-layer contradictions: docs/common, tutorials, and skills teaching
  constructs the validators reject — verify against validation.test.ts and the
  target-language matrix, and check the critic skills especially (their errors
  get amplified by automation).

Audit values

• Name the authority direction for every disagreement before deciding which side
  changes.
• Treat source, tests, and emitted output as evidence; another document's claim
  is only a lead.
• Use independent passes to raise confidence, never as a substitute for cited
  evidence.
• Preserve point-in-time records under docs/history/; correct live docs or
  implementation according to the task and the authority split.
• Make load-bearing enumerable facts mechanically checkable when practical.
  Historical audit precedents live at docs/history/cfc-spec-audit.md and
  docs/history/cts-docs-audit-2026-07.md.