design: TDDs 0044-0047 — coverage map (FR-78) + review/build/tdd-author refinements

[cahenesy]

Four draft TDDs from one /tdd-author pass after the FR-78 PRD merge (PRD-rev
0aa1e28). One covers the new requirement; three are gap-closure refinements of
the how of existing FRs (no PRD change), like 0038/0043.

• 0044 Per-requirement test-coverage map (FR-78): the final review pass emits a
  COVERAGE_MAP block (pinned/proposed/justified-no-surface/unverified-gap)
  ALONGSIDE — outside — its REVIEW_RESULT verdict, so it is structurally
  reported-not-gated (ADR 0005); the runner writes a "## Per-requirement
  coverage" section to report.md + a PR-comment pointer. Anti-false-green is
  model-independent: coverage_map_normalize downgrades a pinned row to
  unverified-gap unless its evidence both matches a ::/:
  citation shape AND the cited file is in the runner-derived git diff --name-only list (ADR 0006 grounding).
• 0045 Review-gate lenses (FR-10, FR-15): intent-conformance
  (documented-but-unenforced is a finding; cite both sides; scoped to the TDD's
  in-scope FR/ADR set) + policy-shadow (a test of an extracted helper is not
  proof the framework calls it; no finding without a named real enforcement
  path). Lenses WITHIN gate 4, not a new gate (ADR 0005).
• 0046 Surgical-changes build norm (FR-66, FR-74): every changed line traces to
  the requirement; no adjacent improvements; minimum code — with a carve-out
  that exempts the build's already-mandated same-commit stale-doc updates,
  failing-test-first commits, and ADR supersessions. Echoed on the rework path
  (the ADR-0008 Opus-rework wander risk).
• 0047 tdd-author interrogation upgrades (FR-76): red-team ranking of
  load-bearing assumptions (impact x likelihood x cheapness, each "fails if
  ___") + a Real/Overblown/Unspoken(elephant) failure-mode taxonomy (recommended
  structure within the existing section, not a new linter heading).

T5 (prd-author objective-level success-metrics) was considered and DROPPED as
redundant with the per-requirement observable acceptance criterion + the FR-77
co-created evaluation rubric (YAGNI).

All four carry an identical co-created ## Evaluation rubric (FR-77). Structural
lint + scope-bound pre-passes exit 0. Also checks off a stale BLOCKERS.md 0038
rework-budget-exhausted entry (moot — 0038 landed; resumed review passed without
rework).

Design-critique gate (independent design-reviewer, fresh context, different
model): DESIGN_REVIEW: PASS. One minor finding addressed before opening this PR —
0044's coverage_map_normalize now carries the scoped-diff file-presence check
into the function spec + verification plan (previously shape-only), making the
stronger anti-false-green half model-independent. Reviewer nits (rubric criterion
wording is 0044-specific but identical per FR-77; NFR-3 noted not-in-scope for
0045) accepted as-is.

Cascade: none stale — all three refinement TDDs are additive gap-closure; the
existing implemented TDDs remain accurate build records.

Open assumptions & waivers:

• FR-78 model false-green on pinned — resolved: pinned requires a cited
  asserting test (file::name or file:line) reproducible from the diff; the runner
  also re-checks the cited file is in the scoped diff, else downgrade. Falsifiable
  per ADR 0006.
• T5 prd-author success-metrics redundant — waived: T5 dropped entirely (YAGNI);
  per-requirement acceptance + the FR-77 rubric already cover measurable success.
• T2 intent-conformance over-blocks — resolved: lens scoped to the TDD's in-scope
  FR/ADR set only; severity by boundary-crossing; no new gate (ADR 0005).
• T2 policy-shadow false positive — resolved: a finding must name the real
  enforcement path (file:line) and show the test misses it; no concrete gap → no
  finding.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com