fix(compat): harden legacy purchase continuation recovery

[bokelley]

Summary

• add revision-fenced, principal-scoped recovery for abandoned legacy purchase operations
• make reservation/completion cancellation-safe and run synchronous executor/reconciler callbacks off the event loop
• validate exact 2.5/3.0/3.1 success, error, submitted, working, and input-required results before durable replay
• bind the complete buyer-visible product/pricing projection and actual peer replay guarantee into continuations
• harden SQLite parent/path/sidecar ownership, CAS transitions, thread snapshots, migrations, and purge locking

Root cause

The initial continuation ledger could durably reserve a seller mutation without leaving a public, fenced recovery path. It also treated arbitrary executor mappings as success, retained only seller-observed pricing rather than the buyer-visible projection, inferred replay safety too broadly, and reopened SQLite through a pathname whose parent was not validated.

Migration behavior

Pre-hardening ledgers migrate in place. Missing execution snapshots can be adopted only by an exact retry, which increments the operation revision. Because old ledgers did not retain the buyer-visible pricing subset or a verified 3.x replay guarantee, unresolved rows remain non-executable; already-terminal results still replay and authoritative reconciliation can still recover an applied result.

Validation

• uv run pytest tests/ -q — 6,835 passed, 40 skipped, 9 deselected, 1 xfailed
• focused continuation suite — 79 passed
• Black, Ruff, mypy, adopter type fixtures, Bandit, and repository pre-commit hooks passed
• independent protocol review: release-ready, no findings
• independent security review: release-ready, no blocking findings

Follow-up to #1059.

[aao-secretariat]

Ladon verdict: Approve

Approve — hardening follow-up to #1059 on the legacy purchase continuation coordinator; no blocking findings.

Checked against the decision table top-down:

• No no-auto-approve team match (HARD RULE clear).
• gated_paths: false, so row 2 does not apply despite review_decision REVIEW_REQUIRED.
• high_risk: true, but both reasons are (modified) files with no medium-or-higher finding — no escalation (rows 3–5 do not fire).
• No prior decision; no sticky escalation.
• Zero medium findings, so row 8 does not fire.
• Falls through to row 9 → approve.

What the reviewer verified: revision-fenced principal-scoped recovery API, cancellation-safe reservation/completion, off-loop sync callbacks, exact source-version result validation before durable replay, buyer-visible pricing projection binding, and SQLite parent/path/ownership/purge hardening. CAS/migration/recovery flows consistent against base and covered by 79 continuation tests. The one semver-prefix note is low and defused (module unreleased).

Nits (non-blocking):

• src/adcp/compat/purchase_continuation.py:476 — new required params added to public exports without ! marker (module unreleased, no live break).
• Migration fail-closed gate and lexical-path symlink rejection confirmed correct.