fix(cliproxyapi): preserve recovery bodies and tool identity

[diegosouzapw]

CLIProxyAPI could return the rejected original request body after a successful retry, lose private tool identity metadata while copying request bodies, and bypass known-field HTTP 400 recovery or repeat already learned reasoning rejections.

The executor now adopts the effective recovery body, composes the existing bounded field-downgrade helper, and applies observed reasoning limits on later requests. Body copies preserve tool-map descriptors; cloaking copies aliases without mutating the caller, and both private metadata channels remain excluded from every HTTP attempt. Shared recovery helpers remain unchanged.

Refs #14629

This is the CLIProxyAPI portion of the umbrella. NineRouter #16133, GLM #16135 and CommandCode #16136 are separate deliveries. Override-contract documentation and the preventive guard remain tracked; this PR does not automatically close the umbrella.

Validation

Three TDD stages on base 122b43b35dbacab03ea67d0f0b46e713280ccd86, each with an identical fixture hash within its failing/passing pair:

• Effective body and private maps: 1 passing control / 5 failures, then 6/6 passed.
• Direct and chained known-field 400: 6 controls / 4 failures, then 10/10 passed.
• Learned capabilities: 10 controls / 5 failures, then 15/15 passed.

The final focused matrix passed 28/28. It exercises both CPA transports, real Responses translation into CPA and restoration of tool aliases/namespaces, caller immutability, retry bounds, provider/model separation, SSE, signal, headers and credential attribution.

All 37 importer/related files passed: 506/506 tests, with zero failures, skipped or cancelled cases. The 28 focused cases are included in this count. No full-repository suite or live CPA instance is claimed.

Core/dashboard typechecks, both complexity gates, focused ESLint/Prettier, file-size, changelog integrity and test discovery passed. Independent SPEC/Standards review found no blockers and verified excluded dependencies unchanged. Tests use synthetic state and intercepted transport.

Normal automatic commit hooks passed with all three hashes unchanged. Source/test, anti-masking and PR-evidence policies passed against the actual committed diff before publication. Hosted CI and merge remain pending.

The anti-masking gate emits one report-only similarity advisory for descriptor?.value instanceof Map. Review confirmed this is a positive shape assertion on the real executor/translator result, followed by descriptor, alias and round-trip assertions; it does not reimplement the copy helper. No suppression or policy change was added.

[github-actions]

CI Coverage Report

• Coverage job: skipped
• PR test policy: success

Coverage artifact was not available for this run.