Signal postmortem FMP migration (#495)

This is the signal-postmortem slice of Batch D. Batches A/B/C and pair-trade are already migrated; #495 stays open for the other skills. The plan was independently reviewed twice. The final plan includes lazy dependency loading and batch-scoped rate-limit state with cleanup.

Final implementation plan

  1. Register a target-only generated client extension for the caller’s exact historical date range. Delegate HTTP, retry, timeout and secret redaction to the shared transport. Remove local HTTP and the legacy v3 fallback.
  2. Load the sibling client lazily and isolate its module name from other skills. Preserve manual/help operation without requests, and standalone use outside this repository.
  3. Keep one lazy client per batch so timing and exhausted-limit state survive between signals. Close its session and restore the prior scope on every exit.
  4. Preserve raw price evidence until validation, including malformed/duplicate dates and non-finite values. Keep the supported historical envelopes and exact requested-range filtering. Never turn unavailable evidence into a flat outcome.
  5. Exercise real CLI paths and fake transport: query parameters, retry/exhaustion, invalid data, redaction, cleanup, import isolation and absent dependencies. Run generator drift, skill/provider/CLI checks and standalone package smoke.
  6. Independently review the complete patch before commit and PR; retain #495 as an open umbrella and do not merge or call live providers.

Contract and limits

The recorder retains fetch_price_data(ticker, start_date, end_date, api_key). The target extension returns raw JSON from /stable/historical-price-eod/full; observed_prices validates it before any return or outcome is calculated. List responses and legacy-shaped historical / matching historicalStockList envelopes remain supported, without requesting legacy endpoints. Missing or ambiguous data returns an empty price map. Only one matching symbol envelope is accepted. A partial batch keeps only supported outcomes; an entirely skipped batch exits 1 and may retain its diagnostic summary.

The shared core can emit redacted transport exception details; the adapter’s own failure summary contains no credential, response body or URL. Transport rate-limit state is reused within a batch. Direct wrapper calls own a temporary session. No normalizer silently drops rows, deduplicates or truncates history before validation. Other generated clients must remain byte-identical.

All tests use fictional data and fake transport. Local and hosted CI results are separate evidence and do not establish live provider availability, trading correctness, or investment performance. No brokerage access is involved.