perf(worktrees): preserve no-op projection identity and skip unchanged evidence replay

[Juliusolsson05]

Confirmed problem

At main 5d64184, canonicalizeWorktreeActivity creates new state/context objects for identical catalogs. The renderer treats those references as changes and appends worktree_catalog_reconciled diagnostics for quiet sessions. LiveWorktreeReconciler.observe also replays up to 500 retained records for unrelated transcript batches.

Evidence

A production-module probe returned new projection references for 1000/1000 value-identical results. In a 120-second audit window, eight sessions each received ten reconciliation records with unchanged diagnostic payloads. A synthetic recorded-fixture benchmark measured approximately 2 ms per unrelated batch with 500 retained worktree events, versus approximately 0.003 ms with none. These are isolated costs, not an app-wide latency measurement.

Expected / acceptance

• Identical canonical projections preserve references.
• Unchanged catalog contents, baseline, cwd and retained evidence reuse the previous projection without replay.
• Relevant evidence, catalog expansion/removal/branch changes and authoritative history replacement invalidate correctly.
• Preserve stale-hydration contribution reversal, evidence eviction, failed-load retry, disposal and session cleanup behavior.
• Add meaningful regression tests and reproducible before/after measurements using sanitized fixtures.

Scope is renderer worktree reconciliation and the shared canonicalizer. Main-process index persistence, #763/#762/#767, and A6 issues #802-805 remain separate increments. No lifecycle/readiness or raw terminal input changes.

[Juliusolsson05]

Implemented in #808 with seven identity/invalidation regressions and a checked-in benchmark. Both CI gates now pass. For 200 unrelated batches with 500 retained records, redundant projection identity changes fell from 200 to zero; isolated median reconciliation cost fell from 2.118166 ms to 0.000333 ms. Not an app-wide typing-latency claim. PR remains open pending review and explicit merge confirmation.

[Juliusolsson05]

Review boundary recorded (from independent review of #808): sameCatalog compares head for every checkout, so ANY commit change in a worktree invalidates catalog identity for that cwd and triggers the full bounded retained-evidence replay at the next refresh. This is correct attribution behavior — an actual checkout change is genuinely relevant evidence — but it is a deliberate boundary of this optimization: a busy repo that moves main every few seconds will not see quiet-session replay savings for sessions on that cwd.