Preserve original query timeout before transport retry

[rmcdaniel]

When an admitted workflow query returns a canonical HTTP 504, the default transport currently sends another POST and can hide the original failure behind a distinct query task. Preserve the first typed QueryFailed response for the three existing query timeout reasons on control-plane v2 query routes, including namespaced runtime URLs.

The focused controls exercise the actual Client path with default 1.20 and explicit 1.19 discovery, asserting one admission and the original status/reason/body. Unknown or malformed responses, unrelated routes/statuses and ordinary transport retries retain existing behavior. Local checks: 39 focused tests, Ruff and mypy pass; the twelve canonical-timeout controls fail against normally installed published 2.5.11.

This addresses the caller retry boundary in #143. Query hydration/lease authority remains open there; the missing original Server timing/status contract is separately tracked in Server #433. The original instrumented failure and frozen published cohorts remain unchanged. Full source gates and a separately installed stable candidate/publication remain required.

Refs #143.

[rmcdaniel]

Complete source CI 38343855047 passes at 3256e2a79ede12b9d71f26e50a4a80dc88d4978c. The normally installed compatible stable candidate passes 239 client/retry tests; all twelve canonical-timeout Client regressions fail published 2.5.11. Ruff and mypy pass, and the corpus guard/history fixtures remain unchanged.

Four actual HTTP controls on unchanged immutable Server 2.6.1 / MySQL / Redis pass with default 1.20 and explicit minimum 1.19. Each starts its own synthetic workflow; the isolated control's query/lease limits are 3/8 seconds, independent of every original failed reference and its deadlines. Published 2.5.11 sends three slow-query POSTs and exposes the final failure after 9.66–9.82 seconds. The candidate sends one, exposes the first original query_worker_execution_timeout / 504 body after 3.19 seconds, and preserves that response exactly. Each control also executes an ordinary fast query returning 7. Real leased query identities and typed late-completion fences are retained.

The first observer setup failed before discovery/admission and its stack was immediately retired. Fresh corrected controls use a separate stack. All runtime/control containers, networks and database volumes are now retired, before the original 04:15 cutoff.

This qualifies the caller retry correction. Query hydration/lease authority remains open in #143, and Server #433 owns the actual timing/status boundary. No Source600 retrospective cause, original deadline extension, frozen retuple or new Nexus capability is inferred. Stable #145 and registry verification remain separate gates.