chore(observability): admin metrics & error logging (#5 review)

[DevHDI]

Summary

Addresses 3 observability findings from the Raven's Verdict review on bmad-method-ui#5. All three were UX/visibility issues that presented fabricated or invisible information; the fixes restore honesty without changing behavior.

Findings addressed

#	Severity	Issue	Fix
#17	🟢 Minor	Admin dashboard showed 0% for "Parsing Errors" — a hardcoded zero presented as a real KPI	Return null from getUsageMetrics, render "N/A — Not yet tracked" in the stats card
#16	🟢 Minor	When a detectBmadRepos GraphQL batch failed, the catch handler set every repo in that chunk to false (indistinguishable from confirmed-absent)	Change return type to boolean | null, mark failed batches as null so callers can distinguish "unknown" from "confirmed no"
#15	🟡 Moderate	fetchFileContent wrapped provider.extendBmadDirs(topSegment) in a silent catch {} — the downstream assertSafePath is still the authoritative invariant but logs had zero signal	Replace the silent catch with a console.warn capturing owner/name/segment/error; no behavior change

Commits

• 85c527c fix(admin): show parsingErrorRate as N/A until tracking exists
• 8090e03 fix(actions): represent detectBmadRepos batch failures as null
• 14113bb chore(actions): log extendBmadDirs failures instead of silently catching

Verification

• pnpm tsc --noEmit — clean
• pnpm lint — 0 errors (2 pre-existing warnings unrelated)
• pnpm test — 151/151 passing

Note on #16

The single in-repo consumer (add-repo-dialog.tsx) does bmadResult.data[r.fullName] ?? false at the boundary, so the null values are coerced to false for the existing badge rendering — current behavior is preserved. Surfacing the unknown count to the user (e.g. "BMAD detection failed for 12 repos") was kept out of this PR because the dialog is also modified by PR #14 for the LIMIT_EXCEEDED warning, and the merge story is cleaner if the UI surface comes as a follow-up after both PRs land.

Context

PR 4 of 4 addressing bmad-method-ui#5 review. This completes the upstream fix series.

• PR 1 ✅ security & input hardening (feat(bmad): BMAD V1 compatibility — dynamic output, frontmatter-driven status, epic-folder layout #7, fix(security): authorization & input hardening (#5 review) #14, feat: local folder support + email/password auth #2) — #14
• PR 2 ✅ revalidate route hardening (feat(bmad): support alphanumeric IDs and introduce agile timeline dashboard #6, feat(epics): Kanban board view with side-panel drill-down #8, Migrate User.role from String to Prisma enum #13) — #15
• PR 3 ✅ docker fail-fast & env hygiene (fix(local): scan only BMAD dirs to avoid file count limit #3, Test #11, Test #10) — #16
• PR 4 ✅ observability (chore(observability): admin metrics & error logging (#5 review) #17, fix(infra): docker fail-fast & env hygiene (#5 review) #16, fix(cache): revalidate route hardening & ordering (#5 review) #15) — this PR

Findings not addressed (intentional):

• feat: support chunked epics directory and add repo settings modal #1 (ESM hoisting in create-admin) — false positive: ALLOW_REGISTRATION is read lazily in Better Auth's before hook, not at module init.
• fix(scripts): load .env in create-admin #5 (path traversal in Zod schema) — LocalProvider.assertSafePath already blocks null bytes, Unicode slashes, traversal, absolute paths, and enforces the bmadDirs whitelist.
• chore(license): relicense to MIT #9 (revalidateTag second arg) — Next.js 16.1.6 accepts revalidateTag(tag, profile). Not spurious.
• chore: add CodeRabbit configuration #18 (double admin auth) — intentional defense-in-depth; server actions can be called outside the RSC page context.

Follow-up issues tracked: #12 CSP and #13 role enum.

[DevHDI]

Follow-up: actually surface the unknown state in the UI

Reviewer's finding was correct. Returning null from the server but letting the dialog do ?? false at the boundary meant the data contract was honest but the user still saw a confirmed-absent label on repos we never confirmed. That's only half of finding #16.

Fix in 56567db

In add-repo-dialog.tsx:

• Count null values in bmadResult.data after a successful detection call
• Track via new unknownDetectCount: number state (distinct from the planned detectWarning state in #14 so both surfaces coexist after merge)
• When > 0, render an amber warning next to the "Detecting BMAD files..." status: "BMAD detection failed for N repositor(y/ies) — these may be mis-labelled as 'no BMAD'."
• The repo list itself stays fully usable; only the misleading badge gets a disclaimer.

The repo's hasBmad still coerces null → false at the boundary, so existing sorting/display logic is unchanged (no need to widen GitHubRepo.hasBmad to boolean | null in this PR).

Verification

• pnpm tsc --noEmit — clean
• pnpm lint — 0 errors (2 pre-existing warnings)
• pnpm test — 151/151

[DevHDI]

JSX whitespace fix

Reviewer was right — splitting the word as `repositor` + `{count === 1 ? "y" : "ies"}` rendered with a space in the middle because JSX preserves whitespace between text and expressions.

Fix in 33c5e95 (pushed)

Move the whole word inside the conditional and use explicit `{" "}` to control the space between the count and the noun:

```tsx
BMAD detection failed for {unknownDetectCount}{" "}
{unknownDetectCount === 1 ? "repository" : "repositories"} —
these may be mis-labelled as “no BMAD”.
```

Renders correctly as BMAD detection failed for 3 repositories — ….

Verification

• pnpm tsc --noEmit — clean
• pnpm lint — 0 errors (2 pre-existing warnings)
• pnpm test — 151/151