fix(locomo): bounded retry on invalid reader completions + protocol-pinned reasoning effort

[fazpu]

What

Two answer-stage robustness fixes, both protocol-identity-relevant, driven by measured v5-strong smoke evidence (2026-07-29):

1. Bounded reader retry. The reader step intermittently receives non-JSON completions from openai/gpt-5.6-luna (3/8 questions in one pass, 1/8 in another), permanently zeroing those questions. Now: up to 2 retries (3 attempts total), only for invalid-completion errors (ProviderInvalidResponseError) after at least one tool call. Every attempt counts against the per-question call budget, the global call budget, and the cost ceiling. Provider outages (transport/HTTP errors) are not retried — the exception hierarchy separates them. Per-item reader_attempts and summary total_reader_retries record the mechanism's activity.

2. Reasoning effort pinned in the protocol. Pinning luna to effort none via ambient REMEMBERSTACK_OPENROUTER_REASONING_EFFORT_MAP raised the smoke 1/8 → 3/8 — but ambient env is invisible to the run pin, violating reproducibility. Now answer_agent_reasoning_effort lives on the protocol dataclass (none for full-v5-strong, None/absent for full-v5) and is applied per-request via a new ModelRequest.reasoning_effort field with explicit-set-wins semantics. Engine seats never set the field, so worker behavior under the env map is byte-identical.

Both fields join the protocol fingerprint (prepare + validate + registry pin check). Fingerprints roll for both entries; all prior runs were smoke-tier diagnostics, nothing published is invalidated. Names and CLI keys unchanged; --protocol remains prepare-only.

Evidence

Smoke v5s-glm52 (pre-pin): 1/8, answer_reader failures 3. Smoke v5s-glm52-b (env-pinned): 3/8, failures 1. This PR makes the winning configuration reproducible from the run pin alone and stops the residual ~12% reader-failure bleed before the full benchmark.

Reviews

• Grok-4.5 adversarial review: APPROVE WITH NOTES — verified retry boundedness, single-count budget accounting, exception MRO (outages not retried), model_fields_set sentinel semantics, engine call sites unchanged, fingerprint coverage on both paths. Notes are test-hardening suggestions (fingerprint identity baseline assert, explicit outage-non-retry test, cost-ceiling mid-retry test), non-blocking.

• Claude (final review): verified budget interactions, MRO, fingerprint derivation in prepare and validate, ran the affected suites (93 + 7 = 100 passed).

• I have read and agree to the RememberStack Contributor License Agreement v1.0.

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
src/rememberstack/adapters
openrouter.py
src/rememberstack/model
__init__.py
model_provider.py
Project Total

This report was generated by python-coverage-comment-action