feat(model-info): add per-turn cache diagnostics

[testikun]

Problem

Issue #156 identified that OpenPI only exposed a cumulative cache percentage. A warm prefix becoming cold could not be located to a turn, while treating every cacheRead=0 as an invalidation would create false claims for first turns, tiny prefixes, and best-effort providers.

Value

Extensions can now consume a bounded, provider-aware observation for each assistant turn and correlate cache changes with local lifecycle boundaries without changing the model context or the existing footer metric.

Approach

• Add a pure cache diagnostics tracker with explicit first-turn, cold, warm, partial-hit, miss-after-warm-prefix, and unknown outcomes.
• Require a prior 2,048-token warm prefix and an explicit-prefix provider contract before reporting a miss; OpenAI/Google best-effort and unknown providers remain unknown.
• Fingerprint the actual system prompt and selected tool surface at Pi's before_agent_start seam.
• Record model, thinking, tool surface, system prompt, compaction, and branch boundaries only as correlations. Every record keeps verifiedCause: null.
• Skip failed and cancelled assistant responses so placeholder zero usage cannot create a false cache miss, replace the valid baseline, or consume pending lifecycle correlations.
• Emit observations on a dedicated extension event channel; preserve cumulative cachePercent, TUI state, and model-visible context unchanged.

Related to #156. Does not close the Issue. Remaining #156 work: a provider usage-field matrix; TTL expiry as its own kind only with evidence; at least one real OpenPI / Bare Pi / OMP Session trace replay before any opt-in marker.

Validation

• bun run check — passed.
• bun run test — passed: Node 1248 passed / 1 skipped; Vitest 30 passed.
• Focused cache diagnostics tests — 14 passed, including failed/cancelled response regressions.
• Updated-main validation: bun run check and bun run test passed with Node 22.19.0, including Vitest.
• git diff --check — passed.

No UI change was made, so a UI screenshot is not applicable.

Impact

• User-visible behavior: none; no marker is enabled by default.
• Model-visible context/tools: none.
• Runtime/lifecycle: one bounded observation is emitted after each assistant turn; tracker state resets per Session.
• Persisted config/data: none.
• Compatibility/risk: additive event surface. Provider classification is intentionally conservative and unsupported providers remain unknown.

[tt-a1i]

Approved at exact head 7e283300f61f1e3bd5aac4c934e2eef95912c4c4. I independently reviewed the scoped implementation and boundary conditions, verified the current diff against main, and found no unresolved P0/P1 issues. Required CI checks are green at this head.