M5d: parity-gated off-call quote delivery

[owieschon]

What changed

• add the SQLite-backed quote delivery queue and explicit off-call worker
• render PDF/XLSX, run the document parity guard, and fail closed before delivery
• add the offline Deliverer protocol and SimulatedDeliverer
• enqueue on caller assent without rendering on the call path
• add HMAC-signed quote retrieval URLs that expire with valid_until
• validate SKU_QUOTE_PUBLIC_BASE_URL and serve /quote/{token} only for valid quotes
• add a 50-run, 1–6 line issuance latency baseline and M5 build-journal receipt

Why

M5c ended with proven renderers and a document guard but no live caller. This
closes the delivery half of M5d while preserving the talk-first latency boundary:
assent only freezes, numbers, journals, and enqueues; document work occurs later.

The safety differentiator is executable. A planted altered-price document invokes
the real parity guard once, reaches no deliverer, blocks the queue row, and records
the failure. A separate frozen oracle drives live assent, proves zero in-call
renderer calls, then retrieves the exact PDF through the delivered URL and rejects
tampered and expired tokens.

Validation

• 754 tests collected: 729 passed, 25 skipped
• python3 -m ruff check .
• python3 -m mypy (78 source files)
• frozen fail-closed and live-route worker oracles
• 50-run latency harness plus independent frozen latency oracle, spanning 1–6 lines
• round-trip, noise-resilience, and independent-resolution audits

The final machine gates are green. Dispatch state is intentionally UNVERIFIED
because the route-probe correction landed after the bounded two-pass cold review;
this draft PR review is the remaining human control.

Out of scope

• live email/SMS providers or credentials
• production worker scheduling
• quote revisions (the next sequential dispatch, after this PR is merged)