fix(optimize): scope transcript-derived findings to the selected provider

[AndrewDongminYoo]

Summary

• scanSessions() hardcoded discoverAllSessions('claude'), so every finding it feeds was computed from Claude transcripts regardless of --provider, while the header came from the already-filtered projects. optimize --provider codex printed a Claude read/edit ratio, Claude ghost skills and Claude MCP findings under a Codex header.
• Skip the scan when the filter excludes Claude, and skip the detectors it feeds rather than handing them an empty scan — emptiness reads as "never invoked", so an empty scan reports every skill, agent and command as unused (19 → 38 on my machine).
• projects-derived findings (MCP tool coverage, capability reliability, low-worth sessions, context bloat, outliers, model recommendations) already filter correctly and still run. cacheKey now carries the provider, since the provider decides whether the scan runs at all.

Fixes #1002

Testing

• I have tested this locally against real data (not just unit tests)
• npm test passes
• npm run build succeeds

npm test: 199 files, 2,631 passed, 5 skipped, 0 failed. Two new cases in tests/optimize-fs.test.ts cover the positive (--provider claude still reports read-edit-ratio) and negative (--provider codex reports neither read-edit-ratio nor unused-skills) paths. I verified the negative case fails without the fix:

AssertionError: expected [ 'read-edit-ratio', …(2) ] to not include 'read-edit-ratio'
 ❯ tests/optimize-fs.test.ts:413

Against real data (30-day window, Claude + Codex + Copilot + Antigravity + Gemini + Warp on disk):

BEFORE
  --provider codex    347 sessions   75,929 calls   Health: F (20/100, 14 issues)
                      ─── 1. Claude edits more than it reads (16,368 reads / 10,510 edits)
                      ─── 10. 19 skills you never use

AFTER
  --provider codex    347 sessions   75,929 calls   Health: D (48/100, 4 issues)
                      ─── 1. 6 MCP servers correlate with retry-heavy edits
                      ─── 2. 79 possibly low-worth expensive sessions
                      ─── 3. 62 context-heavy sessions
                      ─── 4. 2 high-cost session outliers

  --provider claude   2,051 sessions   57,812 calls   Health: F (20/100, 14 issues)   unchanged
  (no flag)           2,409 sessions  134,068 calls   Health: F (20/100, 14 issues)   unchanged

The four findings that survive under --provider codex are genuinely Codex-derived; the Claude-only ones are gone. The default and --provider claude paths are byte-for-byte unchanged in finding count and health grade.