fix(crosscheck): use provider-valid verdict repair effort

[ruby-dlee]

Intent

Fix the released PR 333 verdict-repair incompatibility that blocks PR 327's exact-head Azure Crosscheck. The exact live failure was at PR 327 head 85f89f8: the primary model review completed at xhigh, then the one bounded verdict repair sent reasoning_effort=minimal and direct Fireworks rejected that literal because it accepts low, medium, high, xhigh, max, none, adaptive, integer, or boolean. Reproduce the failure through deterministic adapter/guest tests, identify one owner of repair effort, and make the smallest provider-valid correction. Do not lower the primary xhigh review, weaken schema or evidence gates, add retries, change PR 327 source, touch cloud-result-return files, or invoke another paid review of PR 327. Preserve exactly one bounded fresh repair and terminal telemetry. Add a regression that fails on literal minimal and executes the production command-building/parser path. Validate focused Crosscheck and Azure model-compartment tests, lint, compilation, and docs, then open a green PR and obtain exact-head Crosscheck for that new PR. PR 327's owner will restart its already-authorized review only after this correction lands and deploys.

What Changed

• Use the provider-supported low reasoning effort for the single bounded Pi verdict-repair attempt while preserving the primary review effort.
• Extend the Azure Crosscheck adapter regression to reject unsupported effort values and verify the production repair command and prompt use low instead of minimal.
• Update Crosscheck documentation to describe the low-reasoning repair attempt.

Risk Assessment

✅ Low: Captain, the change is narrowly scoped, preserves the primary xhigh review and single bounded repair, and exercises the production command-building/parser path with provider-valid low effort.

Testing

The previously successful baseline plus the focused Azure Crosscheck suite and a deterministic production-path replay all passed; reviewer-visible evidence confirms xhigh primary effort, exactly one low-effort repair, successful verdict parsing, and preserved terminal telemetry.

Evidence: Deterministic end-to-end verdict-repair transcript

{
"user_flow": "xhigh primary protocol miss -> one fresh low repair -> accepted verdict",
"exit_code": 0,
"attempts": 2,
"thinking_efforts": ["xhigh", "low"],
"verdict": {"summary": "clear"},
"telemetry_turns": 2,
"telemetry_tokens": {
"input": 20,
"output": 4,
"cache_read": 8,
"cache_write": 0,
"source": "pi-turn-end-message-usage"
}
}

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

✅ **Review** - passed

✅ No issues found.

✅ **Test** - passed

✅ No issues found.

• if [ "${FM_AZURE_VALIDATION_CELL:-0}" = 1 ]; then exec "$FM_AZURE_VALIDATION_SHARD_BRIDGE" behavior --count "${FM_AZURE_VALIDATION_SHARD_COUNT:-8}"; else exec bin/fm-no-mistakes-test-command.sh; fi
• Baseline configured test command (reported as already successful by the outer executor).
• tests/fm-crosscheck-azure.test.sh
• Executed the regression’s deterministic fake-provider harness through bin/fm-crosscheck-pi-reviewer.py and bin/fm-crosscheck-pi-verdict-extension.ts using terminal-error-then-valid, then emitted the captured command efforts, verdict, and telemetry.
• git status --short to confirm testing left no worktree artifacts.

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.