Stabilize live plugin validation

[dimavrem22]

Executive Summary

This PR makes complete live plugin validation strict, repeatable, and non-blocking for merges.

• Correlates each current call to one record per owner and reads transcripts from the authoritative owner.
• Makes A2A readiness, mock responses, external-event handoff, and hosted post-call SMS settlement deterministic.
• Preserves exact two-way speech, call mode, action, recipient, accepted-message, and single-delivery behavior.

Description

The live voice harness now sweeps matching active calls, snapshots both owners, applies a request-time lower bound, requires one stable current record per owner, and cleans up every fresh matching record. It checks driver-local speech only on the driver record and requires two-way conversation plus agent-local speech on the AUT record. Hosted scenarios use a speech-safe three-word marker while retaining exact action, target, and single accepted-message assertions.

Hosted post-call turns use a default-deny tool map and a durable SMS journal. Successful sends retain their accepted message ID so later delivery-failure events cannot escape into an unrelated recovery turn and create a duplicate. Pre-delivery policy audit rows remain visible in diagnostics but do not count as accepted SMS messages. A2A receivers are enabled and verified before each scenario, mock completions receive request-unique response IDs, external events acknowledge after durable handoff, setup-only network operations use bounded retries, and public failure output contains only state and counts.

Reason

Live failures were recurring across mutable receiver state, stale reply/call correlation, incorrect transcript ownership, reused mock response IDs, long-running external acknowledgements, model-sensitive hosted actions, and duplicate hosted delivery recovery. The fixes address those failure modes without retrying calls or messages or weakening end-to-end assertions.

Decisions

• Call correlation: Use strict plugin-side snapshots, request-time bounds, directions, numbers, timestamps, and stability because the existing records provide sufficient ownership evidence.
• Hosted settlement: Persist the exact accepted SMS message ID and suppress only its generic delivery recovery so unrelated message failures retain normal behavior.
• Test oracle: Count every fresh accepted marker message while excluding only pre-delivery policy audit rows, which never reached delivery.
• Retry boundary: Retry only idempotent setup and read operations; outbound messages, call placement, and behavioral attempts remain single-shot.
• Merge gating: Keep unit, current-host contract, and loader checks required while the complete live matrix continues to run as non-blocking validation.

Testing

• npm run lint — passed across 152 files.
• npm run typecheck and npm run build — passed.
• npm test — 70 files and 827 tests passed.
• python3 -m unittest tests/live/test_a2a_preflight.py — 3 tests passed.
• Python helper compilation, shell syntax, workflow YAML parsing, and git diff --check — passed.
• Complete live cycle 1 — channels, A2A, voice, external events, and aggregate passed.
• Complete live cycle 2 — the same full matrix passed on the unchanged commit.
• Complete live cycle 3 — the same full matrix passed on the unchanged commit for a third consecutive cycle.
• Fast checks attempts 1–3 — unit, current-host contract, and loader smoke passed in all three attempts.