Add exponential+jitter retry helpers and apply to transient LLM/agent/fs/observation operations

[danielbdyer]

Motivation

• Transient failures (network hiccups, ENOENT races, temporary Playwright navigation errors) should be retried with bounded exponential backoff and jitter while deterministic validation errors must never be retried.
• Centralize retry policy to keep retry semantics consistent and emit retry metadata into rationale/observations for provenance and debugging.

Description

• Added reusable retry primitives in lib/application/effect.ts: exponentialJitterSchedule, a pipeable retryWithBackoff(config), and retryWithBackoffResult(effectFactory, config) which returns retry-attempt metadata (RetryResult).
• Wrapped translation LLM calls (lib/application/translation-provider.ts for llm-api and copilot) with retryWithBackoffResult, gated by a transient error predicate that excludes validation-error, and appended retry counts into translation rationale via withRetryMetadataRationale.
• Wrapped agent interpreter LLM/session calls (lib/application/agent-interpreter-provider.ts) with retryWithBackoffResult, excluded validation-error from retries, and propagated retry metadata into both rationale and observation.detail.retryAttempts using addRetryDetail.
• Added selective filesystem race retries in lib/infrastructure/fs/local-fs.ts (read/write/json/stat/list/ensure) by wrapping tryFileSystem effects with a file-race predicate (checks FileSystemError.cause for ENOENT/EEXIST) and a small backoff.
• Added optional transient retry wrapping to Playwright screen observation (lib/infrastructure/observation/playwright-screen-observer.ts) using a predicate that excludes validation errors and inspects the cause message for navigation/timeouts.
• Retry defaults used at call sites: translation providers baseDelayMs: 150, maxRecurs: 2; agent providers baseDelayMs: 200, maxRecurs: 2; filesystem races baseDelayMs: 40, maxRecurs: 2; Playwright observation baseDelayMs: 120, maxRecurs: 1.

Testing

• Ran npm run build (and npm run types) to verify compilation; the build failed due to pre-existing unrelated TypeScript issues in lib/application/benchmark.ts and lib/playwright/state-topology.ts and not due to the retry changes.
• Verified locally that retry helpers type-check and that no additional compile errors were introduced by the retry changes beyond the unrelated failures reported above.
• No runtime test suite changes were executed in this change; further integration testing is recommended for LLM/agent backends and concurrent FS scenarios to validate retry behavior and rationale/observation metadata emission.

---

Codex Task