fix(sqlite): report lock contention as retryable

[chuiwenwei]

Summary

• classify SQLite SQLITE_BUSY / SQLITE_LOCKED failures as DatabaseBusyError
• return 503 DATABASE_BUSY with Retry-After: 1
• leave unrelated OperationalErrors unchanged
• cover the lock → persistence translation → HTTP response path through the
  production handler set

Closes #498.

Automatic retry is intentionally avoided in the generic session_scope:
replaying an arbitrary transaction after the context body has executed could
duplicate caller-side effects. This change instead exposes an explicit
retryable API contract.

Area

• Architecture method
• Benchmark
• Use case
• Documentation
• Developer experience
• CI, build, or release

Verification

make lint        passed
make test        2704 passed, 4 skipped
make integration 187 passed, 7 skipped, 7 deselected
make package     sdist + wheel built; wheel import smoke passed
make docs-check  passed

Additional regression coverage:

• deterministic lock contention using BEGIN IMMEDIATE
  (tests/integration/test_api/test_sqlite_busy_retry.py)
• unrelated OperationalError remains unchanged
• API response is 503 DATABASE_BUSY with Retry-After: 1
• the same HTTP request succeeds after the lock is released

Checklist

• I kept the change scoped to the relevant area.
• I am opening this from a separate branch, not pushing directly to main.
• I updated docs, examples, or setup notes when behavior changed.
• I added or updated tests when the change affects behavior.
• I did not commit secrets, .env files, dependency folders, or generated output.
• Active relative links in Markdown files resolve.

Notes for Reviewers

The classifier prefers SQLite's primary result codes and retains message
matching as a fallback for wrappers that do not expose sqlite_errorcode.

Please pay particular attention to whether DATABASE_BUSY should remain a
dedicated public error code or reuse the generic infrastructure error code.

• The end-to-end test uses a minimal app with the production handler set,
  not the full /flush route; /flush reaches the same session_scope.
• DatabaseBusyError is a StorageError today. If you want the cascade
  worker to retry it inline (like VectorStoreBusyError), moving it under
  ExternalServiceError or adding a retryable-storage branch is a small
  follow-up — happy to do it.
• The OME's raw aiosqlite store is a separate path and is not covered here.
• The template has no "core runtime" area; I marked Architecture method
  following fix(reflection): batch source deprecation, revert it on failure, cap sources per merge #445.

By submitting this pull request, I agree that my contribution is licensed under
the Apache License 2.0.