fix(v1): preserve cache usage in train responses

[hallerite]

Stacked on the compact-eval-overrides branch to preserve the frozen research base.

Change

Consume optional generate-endpoint usage through CompletionUsage and the existing
Usage.from_openai conversion. This keeps cached prompt tokens out of Usage.prompt_tokens and
preserves optional reasoning details. Responses without usage still use exact token lengths.
The existing serialize_completion already forwards those details to the harness SDK.

Requires renderers' companion fix/forward-generate-usage change for live metadata forwarding.
Without it the existing exact-length fallback remains. No sampling, rendering, graph, numerical
budget, or optimizer changes. The server already reports usage; no serving restart is needed.

Validation

• 84 deterministic verifier tests passed; live-provider/sandbox e2e cases excluded.
• Six model-free cases cover unknown/zero/partial/full cache, reasoning subsets, and missing usage.
• Two tiny live synthetic requests through the actual TrainClient and SDK roundtrip passed.
  Repeated 1,045-token prompt: 1,040 cached + 5 uncached; with 3 output, charge is 8 new tokens.
• Changed-file hooks passed. All-files markdownlint/format passed; unrelated pre-existing E402
  remains in verifiers/v1/harnesses/hermes_agent/program.py:16.
• Type checking was attempted but ty is not installed in the active research environment.

Isolated commits are frozen for fresh local canaries. Existing full comparisons still use their
original dependencies; no result is silently resumed under the fix. Unsigned commit.

Note

Fix response_from_generate to preserve provider CompletionUsage in train responses

• When the generate result includes usage, the parser now validates it as CompletionUsage and converts it via Usage.from_openai, preserving cache and reasoning-token fields from the provider.
• Falls back to the existing prompt/completion token-id length calculation when usage is absent.
• Risk: responses with malformed usage objects may now fail validation instead of silently falling back.

^{Macroscope summarized e4f51c0.}