fix(memory-helpers): handle gbrain >=0.25 doctor non-zero exit and schema_version 2

[mvanhorn]

Summary

/sync-gbrain and every caller of detectEngineTier() now resolve the engine correctly when gbrain doctor exits non-zero (every Supabase brain with health_score < 100) or returns schema_version 2 output (any gbrain >=0.25). Previously these paths collapsed to engine: "unknown" and the orchestrator silently skipped all sync stages.

Why this matters

Two layered bugs both fired for the same users:

1. execSync threw on non-zero exit. gbrain doctor exits 1 whenever health_score < 100 (a fresh Supabase install always has at least one warning, so this is essentially everyone), so the JSON written to stdout never reached the parser. See lib/gstack-memory-helpers.ts:250 (pre-fix).
2. gbrain >=0.25 switched to schema_version: 2 and dropped the top-level engine field. parsed?.engine became undefined, falling straight to "unknown". The project recognizes the new schema elsewhere (test/skill-e2e-brain-privacy-gate.test.ts:52 already stubs {"status":"ok","schema_version":2}) but freshDetectEngineTier was missed during the gbrain >=0.25 update wave.

The v1.29.0.0 release note explicitly bumped the required gbrain version to v0.30.0+, so schema_version 2 support is a project requirement, not a nice-to-have.

Changes

• freshDetectEngineTier calls spawnSync('gbrain', ['doctor', '--json', '--fast'], { stdio: ['ignore', 'pipe', 'ignore'], timeout: 5000 }). The array form drops the shell wrapper while preserving the original 2>/dev/null stderr suppression. Other gstack call sites use the same shape.
• result.stdout is parsed even when result.status is non-zero. Only spawn errors and unparseable JSON fall to "unknown".
• A new detectEngineFromDoctorJson helper implements a defensive lookup ladder: schema_v1 (parsed.engine) first, then schema_v2 top-level renames (engine_kind, backend, engine_type), then a checks[] fallback whose name matches /engine|backend|kind/i reading value/engine/kind. The ladder errs toward "engine resolves" rather than dropping to unknown when one path exists.
• GSTACK_DEBUG-gated stderr warning fires when JSON parses but no engine resolves, so the lookup ladder can be tightened against real-world schema_v2 output. Default behavior is unchanged.

EngineDetect schema is unchanged (schema_version: 1 from gstack's side); the on-disk cache contract is preserved.

Testing

Three new cases under describe("detectEngineTier") cover schema_v1 (existing behavior), schema_v2 healthy, and schema_v2 unhealthy with non-zero exit. The fixtures use a PATH-shim that writes a gbrain bash stub to a tmpdir, prepended in beforeEach and restored in afterEach so the shim does not leak into the cache tests at lines 296 and 305.

$ bun test test/gstack-memory-helpers.test.ts
 25 pass
 0 fail
 59 expect() calls

Fixes #1415

---

^{Need help on this PR? Tag @codesmith with what you need.}

• Let Codesmith autofix CI failures and bot reviews

[mvanhorn]

Heads up @garrytan: the report job is failing on Post PR comment with GraphQL: Resource not accessible by integration (addComment). The workflow's GITHUB_TOKEN is missing pull_requests: write. Adding

permissions:
  contents: read
  pull-requests: write

to the report job (or the workflow) should fix it. PR body itself is fine, CI evals all pass otherwise.