test(model-info): cover cache usage observation boundaries

[ooiuuii]

Problem

Refs #156; this does not close the umbrella issue. The diagnostic core already shipped in #372. Remaining evidence needs to distinguish provider field normalization, OpenPI's existing observational contract, and genuine multi-product Session experiments.

Value

Make the current cache observations reviewable and regression-tested without presenting cache misses as verified TTL expiry, billing evidence, or a proven reason for cache invalidation.

Approach

• Add a source-pinned matrix for Pi 0.85.1 (d981de1229ef899957bbe968bc8dcda02a21f477) and a governed research record distinguishing facts, synthetic evidence, and unfinished experiments.
• Add 16 deterministic cases to the existing model-info tests: implicit/unknown provider boundaries, exact warm-prefix threshold, normalized prompt arithmetic, malformed counters, consumed correlations, and Session restart versus persisted usage.
• Keep all production code, provider classification, configuration, dependencies, and UI unchanged. Tests deliberately exercise the existing implementation; this is not a runtime bug-fix claim.

Validation

Base: OpenPI 5bf2fe29e52801d79826c2eb573be403f53285e5. Candidate: b409b2b3963f2d017864d1e6e55ac0104c4a7860. Local Windows, Node 24.19.0, Bun 1.3.14, frozen lockfile.

Proof: deterministic observation contract

node --test --experimental-strip-types tests/extensions/model-info/cache-diagnostics.test.ts tests/extensions/model-info/index.test.ts tests/extensions/model-info/session-metrics.test.ts
bun run check
bun run test

• Focused model-info tests: 39/39 passed, including 23 pre-existing and 16 added. The checked-in fixtures are small, hand-authored vectors using the actual tracker and extension event harness.
• bun run check: passed (format, lint, typecheck, configuration/discipline checks, Web build).
• git diff --check and new local documentation links: passed.
• Full bun run test: failed locally, not claimed green. Initial Node group: 1,430 passed, 4 failed, 10 skipped. The four failures were in unchanged git-info/process, setup/integration, shared/setup-config, and workflows/narrator tests (timeouts/Windows file-watcher assertion). The same four files passed 41/41 when rerun serially; that does not establish a unique cause or erase the full-run failure.
• Separately executed remaining Vitest group: 9 files / 130 tests passed. The Windows background-terminal serial group had 7 observed failures and then hung; it was stopped after verifying its test-process identity. Its remainder is unverified, not a passing full suite. No unrelated process-lifecycle fixes were folded into this PR.
• Scoped Codex review via autoreview --mode local --engine codex --prompt-file <frozen-scope>: clean, no actionable findings. An earlier attempt ended on a usage limit without a verdict; the completed retry reviewed the same four-file candidate.

Limits / still open

No provider calls, real OpenPI/Bare Pi/OMP Session recordings, cost/latency experiment, false-positive-rate measurement, or verified TTL cause is represented by these fixtures. Those #156 acceptance items remain open. No user chats, credentials, or raw Session files are published. Upstream CI remains the separate repository-wide validation gate.

Impact

• User-visible behavior: none.
• Model-visible context/tools: none.
• Runtime/lifecycle and persisted config/data: unchanged; existing behavior is only documented and tested.
• Compatibility/risk: no peer/lock changes and no new dependencies. This is an independently useful test/documentation slice based on main, with no dependency on the sidebar PR. It intentionally preserves observability: 从累计 cache 命中率升级到逐 turn 的 prompt-cache 失效诊断 #156's remaining evidence boundary.

[ooiuuii]

CI receipt for b409b2b3963f2d017864d1e6e55ac0104c4a7860 (run 34587171375):

• Linux Node 22.19.0 / 24 / 26 each passed bun run check, the full Node suite (1,541 passed, 0 failed, 1 skipped) and Vitest (130 passed). The new model-info cases passed.
• The separate Web E2E job passed.
• The Windows job passed, including both process lifecycle and full-suite isolation. The CI run is now complete; its overall result is still failure because of the packaging step below.
• All three Linux jobs then failed in Smoke-test packed Web CLI with locked Pi host, at npm install, not in the test suites. npm selected @effect/platform-node@4.0.0-rc.114 and reported ETARGET: No matching version found for effect@^4.0.0-rc.114. Node 24 job log.

This PR changes only the two model-info test files and the research record/index; package.json, lockfile, installation commands, and CI are unchanged. The failure points to floating packed-install dependency resolution outside this contribution. I have not changed the dependency policy, retried the workflow, or represented the overall CI as green. The package-install gate still needs resolution separately.

[tt-a1i]

Reviewed current head 1db7963 after updating to the Effect cohort fix. Standards: no actionable findings. Spec: the added synthetic tests and source-scoped cache usage record distinguish observation from verified causes, and explicitly leave real comparative Session evidence open. This adds no runtime or model behavior. No P0/P1 found. Merge remains conditional on all required current-head CI.