test: add provider usage-normalization, compaction-telemetry, prompt-prefix coverage

[fxinfo24]

Summary

Three test modules that were written against existing behaviour but never committed — so they have not been running in CI. All pass on Python 3.10 and 3.11.

Module	Tests	Covers
tests/providers/test_usage_normalization.py	30	normalizeUsage() across every provider
tests/test_compaction_telemetry.py	17	Cache-hostile compaction detection
tests/prompt/test_prefix_stability.py	15	Prompt-prefix stability

Why the usage-normalization matrix matters

The tests pin each provider's normalizeUsage() to the Anthropic wire convention:

• input_tokens — cache MISS, priced at the input rate
• cache_read_input_tokens — cache HIT, priced at the cache-read rate
• cache_creation_input_tokens — cache WRITE, priced at the cache-write rate

This is the audit that catches an OpenAI-wire-format provider folding prompt_tokens_details.cached_tokens back into prompt_tokens. That mistake is silent — the request still succeeds and token totals still look plausible — but it overstates cache-miss spend, which is precisely the number prompt-caching economics depend on.

One fix beyond adding files

TestOpenAICompatibleProvider is a helper base class, not a test class. pytest collects by the Test prefix, so it emitted a PytestCollectionWarning on every run (it also inherits an __init__ from MockMessage).

Marked it __test__ = False. Nothing subclasses it, so this suppresses no real tests — verified by confirming zero subclasses before applying.

Verification

• These three files: 62 passed on both Python 3.11 (CI's pinned version) and 3.10 (the declared floor).
• Full suite on this branch: 10889 passed, 15 skipped, 4 failed.

The 4 full-suite failures are pre-existing, not from this PR

Confirmed by running the same tests on an untouched fork/main checkout.

Three failures in the full-suite run, all in test_init_integration.py, are the SIGINT/SIGTERM startup race that #961 fixes:

• TestSigtermPathRunsCleanups::test_sigterm_triggers_drain
• TestSigintDuringPrefetch::test_sigint_during_prefetch_clean_exit
• TestSigintBeforePrefetchStarted::test_sigint_before_any_prefetch

(The third passes in some runs — the race is timing-dependent, so it appears when that file is run on its own more reliably than in a full-suite run. Same root cause.)

Two further failures reproduce on pristine fork/main with none of my changes applied, and are unrelated to this branch:

Failing test	Cause
workflow/test_agent_concurrency.py::test_concurrent_agents_run_in_parallel	Timing assertion (9.76s < 1.0s)
bridge/test_bridge_main.py::test_spawn_worktree_creates_isolated_dir_and_cleans_up	Spawner race (assert 0 == 1)

Note

No production code is touched here — this is tests only, all passing against current behaviour on a clean checkout.

[github-actions]

Test Results

     5 files   1 021 suites   21m 3s ⏱️
16 166 tests 16 144 ✅ 22 💤 0 ❌
32 302 runs  32 231 ✅ 71 💤 0 ❌

Results for commit 26f7737.