v0.25.0 - Trustworthy Adaptive Retrieval

v0.25.0 - Trustworthy Adaptive Retrieval

Agent Memory Bridge 0.25.0 makes retrieval behavior easier to inspect without
turning feedback into authority. The public MCP surface grows to 13 tools with
receipt-bound feedback.

What changed

• Retrieval capability is labeled as lexical, hashed_lexical, or
  semantic. The default hash path remains hashed_lexical; true semantic
  mode requires a declared semantic provider.
• Explicit non-empty kind="memory" text recall can return a short-lived HMAC
  receipt bound to bridge instance, database epoch, namespace, query hash,
  retrieval mode, and memory ids/ranks.
• feedback records a declared outcome for one recalled result:
  helpful, misleading, outdated, not_applicable, or not_used.
  misleading and outdated require a compact reason.
• Feedback is append-only, idempotent for stable retry identity, redacted in
  responses/logs/telemetry, and shadow-only.
• First-run and verify guidance now use the v0.25.0 install tag.

Validation snapshot

• pytest --collect-only -q tests: 604 tests collected
• Full suite result: 601 pass, 3 skip
• Public MCP surface: exactly 13 tools
• Release, public-surface, onboarding, stdio feedback, v0.25 receipt, v0.25
  feedback, v0.25 retrieval-capability, and v0.25 migration tests are the main
  release-surface gates for this patch.

Boundaries

Feedback does not mutate memory records, FTS rows, embedding rows, belief
records, returned ordering, or ranking behavior. Receipt tokens are
tamper-evident, not encrypted. Caller/provenance fields are declared, not
authenticated. The database epoch is a restore-instance guard for stale
receipts and cursors, not per-write freshness.

This release does not add active reranking, authenticated actors, namespace
ACLs, ANN retrieval, graph memory, automatic policy, hosted identity, or
multi-user infrastructure.