OmniMemory

1. Executive Summary

OmniMemory is the memory domain package of the ONEX platform: typed models, Kafka-dispatched nodes and storage adapters behind a kernel plugin. What is notable is a test file that tests its own checker, and an expire path that reports an outcome for every command. What is weak is that the routed retrieval node can only serve in-memory test doubles nothing seeds, so the running system returns nothing from memory.

The thing to read first is a test file.

tests/gates/test_consumer_group_name_authorization.py enforces that no Kafka consumer-group name in src/ is a string literal. It is a default-deny AST walk: every group_id= / consumer_group= / kafka_group_id= keyword and every default on a matching field "must be a call to (or a local binding of) one of the canonical derivation helpers. A newly added string literal FAILS by default -- that is the whole point of the gate."

Then it tests the checker:

"A default-deny gate whose discrimination is never exercised is indistinguishable from a gate that always passes; these prove it catches literals, f-strings, and laundering while admitting real derivations."

Five test_checker_* cases run positive and negative controls over the AST walk against inline fixtures. The module docstring records why they exist: they were "added in the OMN-15639 round-2 remediation after a verifier demonstrated the apply_instance_discriminator laundering hole live."

test_unauthorized_env_token_is_not_authorized asserts, in executable form, what the suite does not prove. The kernel's default environment token is local, local.omnimemory.* is not in the IAM-granted pattern set, and therefore "a green suite here … does NOT prove the deployed name is authorized; the deployment must set the env token to onex-dev." The test's subject is the boundary of the suite's own evidence.

The second mechanism is a compare-and-set on lifecycle, wired to outcome events. HandlerMemoryExpire transitions a memory with a WHERE clause carrying both the expected revision and the expected source state, and reports zero affected rows as a conflict (handler_memory_expire.py:381-391). create_expire_memory_dispatch_handler routes every expire-memory command to it and publishes exactly one event: memory-expired.v1 on success, lifecycle-transition-failed.v1 on refusal, malformed command, missing database or conflict. With no publisher wired it raises before touching the database, so the message is nacked rather than acknowledged in silence (runtime/dispatch_handlers.py:927-1082).

The finding that governs the report is on the read side. The retrieval dispatcher builds HandlerMemoryRetrieval from build_retrieval_config_from_env(). OMNIMEMORY_USE_STUB_HANDLERS defaults to "true", which selects in-memory mocks that nothing in src seeds; any other value reaches raise NotImplementedError("Production handlers not yet implemented. Set use_stub_handlers=True") at initialisation (handler_memory_retrieval.py:151-170). HandlerQdrant exists, is tested, and has no constructor in src.

The other two lifecycle commands are acknowledged without acting. archive-memory reaches a bridge that logs "archive-memory command not yet implemented … acknowledging as no-op", and runtime-tick reaches the same bridge, which starts adapters and returns.

2. Mental Model

There are two OmniMemories in this repository, and they meet at one point.

The specified one is careful: typed enums, a frozen transition map, a compare-and-set handler with a conflict result and a retry helper, and an archive handler that writes gzip JSONL atomically before it updates the row. It also specifies a promotion tier with a QUARANTINE state "not included in context assembly by default", an audit metadata model with old_value/new_value, and a gate suite that tests itself.

The running one routes expire-memory to the compare-and-set and reports the outcome. It routes archive-memory and runtime-tick to a bridge that does not call the archive or tick handlers. Its retrieval serves unseeded test doubles. It never constructs the audit model, never writes a promotion tier, and never consults the transition map.

A memory is a memories row with a lifecycle_state. It stops being current in exactly one routed way: an external service sends expire-memory naming the row and the revision it last saw, and the row moves ACTIVE → EXPIRED if both still match. Nothing in this repository inserts a memories row, nothing moves one to STALE, ARCHIVED or DELETED on a routed path, and nothing on any read path asks what state a row is in.

%% caption: what is specified, what each lifecycle command reaches, and what the read path serves
flowchart TD
    subgraph SPEC["specified, no routed caller"]
      VT["VALID_TRANSITIONS, frozen<br/>STALE→ACTIVE, ARCHIVED→ACTIVE<br/>DELETED terminal"]
      ARC["HandlerMemoryArchive<br/>gzip JSONL, then CAS to ARCHIVED"]
      TICK["HandlerMemoryTick<br/>active and expires_at ≤ now<br/>expired and archived_at IS NULL"]
      PT["EnumPromotionTier<br/>QUARANTINE · VALIDATED · SHARED · BLACKLISTED"]
      AM["ModelAuditMetadata<br/>old_value, new_value"]
    end
    subgraph RUN["routed by PluginMemory"]
      EXP["expire-memory"] --> ED["create_expire_memory_dispatch_handler"]
      ED --> CAS["UPDATE memories … WHERE id AND revision<br/>AND lifecycle_state = 'active'"]
      CAS -- "row returned" --> OK["publish memory-expired.v1"]
      CAS -- "no row, error, or no handler" --> FAIL["publish lifecycle-transition-failed.v1"]
      ED -- "no publisher" --> NACK["raise: message nacked"]
      AT["archive-memory, runtime-tick"] --> BR["lifecycle bridge:<br/>start adapters, log, return []"]
      RQ["memory-retrieval-requested"] --> RH["HandlerMemoryRetrieval"]
      RH -- "USE_STUB_HANDLERS default true" --> MOCK["in-memory mocks,<br/>never seeded"]
      RH -- "any other value" --> NIE["NotImplementedError"]
    end
    BR -. "never calls" .-> ARC
    BR -. "never calls" .-> TICK
    MOCK -. "no lifecycle_state predicate" .-> CAS

3. Architecture

ONEX-compliant nodes, each with a contract.yaml, dispatched from Kafka topics derived by helper functions rather than literals. PluginMemory registers at the onex.domain_plugins entry point, builds a MessageDispatchEngine in create_memory_dispatch_engine, and subscribes to the topics the contracts declare. The lifecycle orchestrator's contract.yaml routes tick, expire and archive operations to HandlerMemoryTick, HandlerMemoryExpire and HandlerMemoryArchive, but its package is not in _OMNIMEMORY_HANDLER_NODE_PACKAGES (runtime/wiring.py:44-51), and in this repository only the expire handler is constructed, by PluginMemory._build_expire_handler (runtime/plugin.py:466-519).

The README lists node handlers as "contract-carrying nodes in src/omnimemory/nodes/, migrating to omnimarket", and names PostgreSQL and Kafka as owned by omnibase_infra. The boundary document it links lives in a separate knowledge-base repository.

Deployment and ergonomics

docker-compose.yml owns Qdrant, Memgraph, Valkey and a Kreuzberg parser. PostgreSQL and Kafka come from the platform. The expire handler needs OMNIMEMORY_DB_URL; unset or unreachable, the plugin logs a warning and every expire command is answered with lifecycle-transition-failed. There is no single-process mode and nothing an agent installs directly.

4. Essential Implementation Paths

Expire — runtime/plugin.py:711 builds the handler; runtime/dispatch_handlers.py:927 create_expire_memory_dispatch_handler validates the payload as ModelExpireMemoryCommand in lax mode after dropping the envelope keys correlation_id, requested_by and timestamp, calls HandlerMemoryExpire.handle once, and publishes one event. handle_with_retry is not on this path.

Transition — nodes/node_memory_lifecycle_orchestrator/handlers/handler_memory_expire.py (_EXPIRE_SQL :381, _READ_STATE_SQL :394, _VALID_FROM_STATES :404, handle :565). A Postgres error is caught and returned as success=False, error_message="Database error: …", which the dispatcher turns into a failure event.

Validate — validators/validator_lifecycle_transition.py (VALID_TRANSITIONS as a MappingProxyType of frozensets, :63-96). No module outside the validators package imports it.

Archive and sweep — handlers/handler_memory_archive.py and handlers/handler_memory_tick.py (lifecycle_state = 'active' AND expires_at IS NOT NULL AND expires_at <= :now, then lifecycle_state = 'expired' AND archived_at IS NULL). Their only constructors in src are docstring examples and the two adapters in adapters/, which nothing constructs.

Lifecycle bridge — runtime/dispatch_handlers.py:825 _create_lifecycle_bridge_handler: reads payload["command"], starts HandlerMemoryLifecycle if needed, calls handle_shutdown() only for "shutdown", logs and acknowledges "archive-memory".

Retrieve — runtime/dispatch_handlers.py:467 build_retrieval_config_from_env, :515 create_memory_retrieval_dispatch_handler, then nodes/node_memory_retrieval_effect/handlers/handler_memory_retrieval.py (initialize :151-170, _execute_hybrid :347).

5. Memory Data Model

deployment/database/migrations/002_create_memories_table.sql creates memories with content, content_type, lifecycle_state (CHECK over five values, default active), expires_at, archived_at, lifecycle_revision, archive_path, metadata, created_at and updated_at. It has no expired_at column. _EXPIRE_SQL sets expired_at = $2, and HandlerMemoryArchive._read_memory selects it. Against a table built from this migration the UPDATE names a column that does not exist; the handler would return the Postgres error and the dispatcher would publish lifecycle-transition-failed. The README places PostgreSQL ownership in omnibase_infra, so whether the deployed table has the column is outside this tree.

EnumLifecycleState is five states with the transitions written down: STALE can return to ACTIVE ("refreshed / promoted"), ARCHIVED can return to ACTIVE ("promoted / restored"), and DELETED has an empty frozenset, described as "soft delete marker for audit trail". Revival paths named explicitly is the right way to write a lifecycle; no routed code performs any of them.

EnumTrustLevel looks like a trust state and is not one. Its six levels are defined by score ranges — "UNTRUSTED: Score below 0.2, data should not be used" — and model_trust_score.py derives the level from the float in _score_to_level() (:161). It is a confidence number with labels printed on it.

EnumPromotionTier is closer. It is a stored tier "in PostgreSQL and Qdrant payload" with QUARANTINE ("Not included in context assembly by default"), VALIDATED, SHARED and BLACKLISTED ("permanently excluded"), and ModelContextPolicyConfig.include_quarantine defaults to False. The docstring names ContextSelectorNode as the reader. No module in this repository defines that node, reads include_quarantine, or constructs ModelPromotionDecision; the tier has no writer and no reader here.

6. Retrieval Mechanics

The routed retrieval node serves in-memory test doubles. build_retrieval_config_from_env() reads OMNIMEMORY_USE_STUB_HANDLERS with default "true" and, for any other value, builds a ModelHandlerQdrantConfig. HandlerMemoryRetrieval.initialize() then initialises HandlerQdrantMock, HandlerDbMock and HandlerGraphMock in the stub case and raises NotImplementedError in every other case. The mocks start empty; seed_snapshots has no caller in src outside docstrings. A deployment that leaves the default retrieves nothing, and one that sets it to "false" fails on the first request.

HandlerQdrant is a working vector search: it embeds the query and calls qdrant_client.search with limit and score_threshold and no filter (handler_qdrant.py:230-278). Its tests construct it; nothing in src does.

lifecycle_state appears in no file under the retrieval node or runtime/. An EXPIRED or ARCHIVED memory would not be filtered out by any read path found here.

The composed read path

_execute_hybrid fans out to a semantic leg and a full-text leg and fuses their rankings rather than their scores, with the reason written down: fusing first "would let a leg that matched" on a stray term carry a document into the answer on one leg's strength alone. fuse_rrf_scores is reciprocal-rank fusion at RRF_K_DEFAULT = 60 (handler_fusion.py:53), weighted rather than plain, because on a paraphrase query the lexical leg can match a stray token and unweighted RRF would promote it.

handler_decay.py then modulates the fused score multiplicatively — A(t) = A_0 * exp(-lambda * dt) with DECAY_LAMBDA_PER_DAY = 0.015 — and the ordering is deliberate: decay is applied after fusion, "not as a third ranked list inside the fusion."

The freshness input degrades per instance rather than per model: the plan specifies max(created_at, last_accessed_at), but the retrieval path returns a frozen ModelMemorySnapshot from a shared core package that carries only created_at, so freshest_timestamp (:83) falls back per record. The docstring states the policy that follows — whether the created_at-only path costs ranking quality "is a question to answer with the fixture harness, not by assumption."

STALE_ACTIVATION_THRESHOLD = 0.3 (:73) is exported with the stated reason that "ranking and the lifecycle orchestrator must agree on what 'stale' means". is_stale appears in src only at its definition and in the module's __all__; six committed cases are its only callers. Staleness is a defined, tested predicate that nothing filters or expires on.

All of this runs over the mocks. The fusion and decay code is real and tested; the backends it fuses in a deployment are the empty doubles above.

No memory-level scope key — tenant, user, project or agent — is applied as a read filter. AdapterPostgresPersona.get_latest selects by user_id and the subscription handler by agent_id; neither reads memories.

7. Write Mechanics

Nothing in this repository inserts into memories. The memory storage node writes snapshots as JSON files through a filesystem adapter, navigation history writes to navigation_sessions and Qdrant navigation_paths, and persona snapshots are inserted into user_persona_snapshots.

Lifecycle changes go through the compare-and-set, which is a lost-update guard: the WHERE clause carries the expected revision and the expected state, and zero affected rows is reported as conflict rather than swallowed. The expire command supplies expected_revision; the dispatcher makes one attempt and reports a conflict as a failure event with conflict: true and the state it found. _VALID_FROM_STATES = frozenset({EnumLifecycleState.ACTIVE}) carries the comment that keeps it narrow: "Including EXPIRED here would cause handle_with_retry() to keep retrying on already-expired memories."

Operational cost

One UPDATE per expire command, plus one SELECT when it affects no rows. The pool is five connections. Nothing rewrites the store in the background, because the tick handler is not routed.

8. Agent Integration

Not an agent-facing memory in the usual sense: no MCP server in src (the mcp dependency in pyproject.toml is commented as Archon connectivity), no CLI for a coding agent. It is a domain package inside a distributed platform, consumed by other ONEX services over Kafka.

9. Reliability, Safety, and Trust

No marks.

Trust state — withheld. EnumTrustLevel is a binning of a float. EnumPromotionTier has the right shape — a QUARANTINE state withheld from context assembly by default — and no producer or consumer in this tree. EnumLifecycleState is a lifecycle, not an epistemic status, and no read path consults it.

Audit log — withheld. models/foundation/model_audit_metadata.py defines AuditEventDetails with operation_type, resource_id, old_value, new_value, request_parameters, response_status, error_details, ip_address and user_agent, plus ResourceUsageMetadata and SecurityAuditDetails. Only tests/models/test_frozen_boundary_models.py constructs any of them. The memory-expired.v1 event is a bus message, not a record in the system's own store, and no migration here creates an event or history table.

Scope, bitemporal, tombstone, human review, negative eval — no. The nearest negative retrieval case, test_search_no_results, asserts an empty mock store returns nothing; test_select_relevant_preserves_rank_order_and_drops_below_floor asserts a score floor on synthetic entries. Neither asserts that particular stored material must stay out.

The outcome contract is the strongest property. An expire command produces an event or a nack; the fix commit says so and tests/unit/runtime/test_expire_memory_dispatch.py pins it through create_dispatch_callback, the runtime's own entry point. That test file runs HandlerMemoryExpire over _FakePool, whose fetchrow replays scripted rows regardless of the SQL, so the suite cannot see the expired_at column gap.

The archive acknowledgement is the risk to name. An upstream service that sends archive-memory receives the same acknowledgement as a success, and the warning lives in a log line.

10. Tests, Evals, and Benchmarks

No paper. No retrieval benchmark with a live backend. mypy strict, ruff, pre-commit.

tests/gates/test_consumer_group_name_authorization.py, ten tests:

• test_no_unmigrated_group_literals_in_src — the default-deny AST walk.
• test_legacy_unmigrated_allowlist_is_empty — the escape hatch, asserted empty, so the exemption list cannot quietly grow.
• Five test_checker_* cases proving the walk "catches literals, f-strings, and laundering while admitting real derivations", including test_checker_rejects_laundered_literal and test_checker_rejects_rebound_derived_name.
• test_derived_memory_group_is_iam_authorized — positive and negative control against a pinned IAM pattern set vendored from Terraform, because "without the negative half a permissive translator would pass".
• test_seam_scope_field_name_is_ephemeral_tag — pins a field name across a package boundary the seam table does not cover, "so a core-lane rename fails here loudly instead of surfacing as a runtime TypeError while both lanes' own suites stay green".
• test_unauthorized_env_token_is_not_authorized — the test that states the suite's own limit.

tests/unit/runtime/test_expire_memory_dispatch.py drives a raw bus message through the engine and asserts one success event for an ACTIVE row and one failure event, and no success event, for a non-ACTIVE row, an unknown row, a revision conflict, a missing handler and a malformed command. Two more assert a publish failure and a missing publisher each raise. A last one asserts the lifecycle bridge does not read expire-memory as shutdown.

tests/unit/nodes/test_hybrid_retrieval_fusion.py scores fusion against a committed 32-query corpus by nDCG and asserts that unweighted RRF does regress paraphrase queries, so a change that removes the weighting fails.

I ran nothing.

11. For Your Own Build

Steal

• Test the checker, not just the thing checked. "A default-deny gate whose discrimination is never exercised is indistinguishable from a gate that always passes." Every invariant you assert deserves a fixture proving it fails when violated.
• Write a test for what your tests do not prove. Naming the residual in executable form — here, that a green suite says nothing about the deployed environment token — outlasts a paragraph in a PR body, because the paragraph is not run.
• Assert the escape hatch is empty. _LEGACY_UNMIGRATED "MUST stay empty at end state", enforced by its own test. An allowlist nobody checks grows.
• Put both the revision and the source state in the WHERE clause. WHERE id = ? AND lifecycle_revision = ? AND lifecycle_state = 'active', with zero rows surfaced as a conflict, gives you optimistic concurrency and an illegal-transition guard in the same statement.
• Answer every command with exactly one outcome, or refuse to take it. The expire dispatcher publishes success or failure for each command and raises when it has nowhere to publish, so the bus redelivers instead of losing the answer.
• Name the revival paths in the transition map. STALE→ACTIVE and ARCHIVED→ACTIVE spelled out, DELETED with an empty frozenset. Someone reading the map learns the policy.

Avoid

• Do not acknowledge a command you did not perform. archive-memory is logged as "not yet implemented" and acknowledged; the upstream sender cannot tell that from success. The expire dispatcher in the same file shows the alternative.
• Do not let a config switch select between test doubles and an exception. A retrieval node whose default is an unseeded mock and whose alternative is NotImplementedError cannot be configured into working; the failure mode of the default is a memory that silently returns nothing.
• Do not test SQL against a pool that ignores the SQL. A fake whose fetchrow replays rows will pass an UPDATE naming a column the migration never created.
• Do not let a score wear a state's clothes. Six labels derived from a float by threshold are a float. If UNTRUSTED means "should not be used", something on the read path has to refuse it.
• Do not ship an audit schema with no writer. old_value, new_value, ip_address and user_agent, fully specified and never constructed, reads to a reviewer as an audit trail that exists.

Fit

This is a platform component, not a library to adopt. If you are not building on ONEX there is nothing here to install, and if you are, the read side is not finished in this repository.

Read the gate file anyway. Its central claim — that an unexercised gate and an absent gate are the same gate — applies to every invariant a memory system asserts, and the expire dispatcher next to it is a compact example of a command handler that cannot fail silently.

12. Open Questions

• Is the lifecycle wired anywhere outside this repository? The README lists node handlers as migrating to omnimarket, which was not read; the lifecycle orchestrator's contract.yaml declares tick, expire and archive routing that a runtime there might consume.
• Who creates the memories table in a deployment, and does it have expired_at? PostgreSQL is listed as owned by omnibase_infra, which was not read.
• Who sends expire-memory, and from where does it learn expected_revision? No producer is in this tree.
• What does OMN-2437 say? The ticket that turned a fail-fast lifecycle handler into the logged no-op that create_lifecycle_dispatch_handler returns is referenced and not public; that factory has no caller outside tests.
• Is ModelAuditMetadata written, or EnumPromotionTier read, by a sibling package? Neither has a writer or reader here; omnibase_core and the undefined ContextSelectorNode were not read.

Appendix: File Index

Recorded searches

Run at the tree root of the pinned checkout with /usr/bin/git grep.

• git grep -n 'lifecycle_state' -- src/omnimemory/nodes/node_memory_retrieval_effect src/omnimemory/runtime — no match; no read path consults lifecycle.
• git grep -nE 'HandlerQdrant\(' -- src — no match; the Qdrant retrieval handler is never constructed.
• git grep -n 'Production handlers not yet implemented' -- src — handler_memory_retrieval.py:169, the non-stub branch.
• git grep -n 'seed_snapshots(' -- src — docstrings and the handlers' own methods only.
• git grep -nE '(HandlerMemoryArchive|HandlerMemoryTick|HandlerMemoryExpire|AdapterRuntimeTickMemory|AdapterPostgresDeactivateMemory)\(' -- src — docstring examples, the two unconstructed adapters, and runtime/plugin.py:517, the one production constructor.
• git grep -nE 'validate_transition|VALID_TRANSITIONS' -- src ':!src/omnimemory/nodes/node_memory_lifecycle_orchestrator/validators' — no match.
• git grep -n 'expired_at' -- '*.sql' — no match; the migration lacks the column the expire UPDATE writes.
• git grep -niE 'insert into memories' -- . — no match.
• git grep -nE '(AuditMetadata|AuditEventDetails|SecurityAuditDetails|ResourceUsageMetadata)\(' -- . ':!src/omnimemory/models/foundation/model_audit_metadata.py' — four hits, all in tests/models/test_frozen_boundary_models.py.
• git grep -nE 'TrustScore\(' -- src — the class definition only.
• git grep -nw is_stale -- src — the definition and __all__.
• git grep -nw -e include_quarantine -e ContextSelectorNode -- src — the field and a docstring.
• git grep -nE 'ModelPromotionDecision\(' -- src — the class definition only.
• git grep -nE '(WHERE|AND).{0,80}(user_id|tenant_id|agent_id|project_id|session_id|owner_id|namespace|workspace|scope)' -- . ':!uv.lock' — subscriptions by agent_id, persona by user_id, and test SQL on navigation_sessions.
• git grep -nE 'query_filter|FieldCondition|MatchValue' -- src — no match; no Qdrant payload filter.
• git grep -niE 'tombstone|rejected_value|do_not_(store|remember)' -- src deployment tests — no match.
• git grep -nE 'valid_from|valid_to|valid_until|invalid_at|event_time|effective_(from|at)' -- src deployment — valid_from_states only.
• git grep -niE 'pending_review|needs_review|review_status|approv(e|al|ed)' -- src deployment — no match.
• git grep -nliE 'FastMCP|mcp\.tool|mcp\.server|from mcp' -- src — no match.
• git grep -niE 'arxiv|bibtex|@article|@misc|doi\.org' -- README.md CLAUDE.md CHANGELOG.md — no match; no paper.
• git grep -niE 'anthropic|machine learning|not (be )?used (for|to)' -- LICENSE — no match; the MIT text carries no rider.

History

2026-10-12 — 7be7aa9e… — 96 commits on. 7be7aa9 routes expire-memory to HandlerMemoryExpire with one outcome event per command (section 4); the compare-and-set has a production caller. Marks unchanged at none. Four claims were wrong at the previous pin, the first three at every pin: the routed lifecycle handler was a bridge, not the no-op factory; retrieval stubs were the only initialisable path, not a default (section 6); the checker controls number five; the migration-boundary quotation and docs/migrations/ path were gone by 1 September 2026. Newly described: the unwired promotion tier and the missing expired_at column (section 5). Screened from a full clone: no auto-run surfaces, four pytest conftest.py files, pyproject.toml and uv.lock inside the cooldown, CLAUDE.md read as data. Nothing was installed or run.

2026-09-17 — 59874e5b… — re-read twenty commits on and the memory subsystem did not move. Marks unchanged at none. Every file this report anchors is byte-identical by blob sha: both enums, model_audit_metadata.py, model_trust_score.py, contract_topics.py, dispatch_handlers.py and plugin.py. What changed is twelve files of dependency bumps, a commit-msg hook, a CI-alignment test and a lock refresh by a bot; the three source files it touched — handler_subscription.py, adapter_filesystem.py and introspection.py — are none of the ones the findings rest on.

All three absence claims were re-run rather than carried forward, and all three hold. AuditMetadata( has no constructor anywhere in src, so the audit model is still never written. TrustScore( has none either. And lifecycle_state still appears in no retrieval query: its only SQL sites are the lifecycle orchestrator's own expire and tick sweeps, and handler_memory_retrieval.py does not mention it — so an expired or archived memory is still not filtered out on the read path. That is the report's central finding and it is unchanged at this pin. Screened again before reading: no auto-run surface, four build-time execution surfaces, no unpinned surface, one dependency file inside the seven-day cooldown. Nothing was installed and nothing was run.

2026-09-09 — 6d340f89… — second reading, 54 commits along the default dev branch. Screened before reading; nothing was installed and no suite was run.

Every path this report's appendix names is byte-identical to the previous pin — git diff --stat <old>..HEAD -- over the gate, the lifecycle node and the runtime returns nothing, and every cited line anchor still resolves to the symbol it named. The commit subjects are dependency bumps, sibling-lock refreshes, CI gates and doc scrubbing.

An empty appendix diff is weak evidence that nothing happened, and here it was wrong. Three handlers arrived beside the paths the appendix lists — handler_memory_retrieval.py, handler_fusion.py and handler_decay.py, 437 lines together — carrying a composed read path the report did not describe: rank-level reciprocal-rank fusion at k = 60, weighted so a stray lexical match cannot carry a document alone, then a multiplicative activation decay at lambda = 0.015 per day applied after fusion rather than as a third ranked list. Section 6 now covers it.

The most useful thing in it is a refusal. The freshness input degrades per instance rather than per model, because the retrieval path returns a frozen snapshot from a shared core package that carries only created_at, and the module says why it declines to change that model: whether the fallback costs ranking quality "is a question to answer with the fixture harness, not by assumption."

No mark moves, and one near-miss is worth stating. STALE_ACTIVATION_THRESHOLD = 0.3 is exported so that "ranking and the lifecycle orchestrator must agree on what 'stale' means", and is_stale has six committed cases — but rg -n "is_stale" src/ finds the name only in its own __all__. It is a defined, tested predicate with no caller, so nothing filters or expires on it and trust_state stays withheld.

Tests grew by roughly 1,700 lines across activation decay, decay ranking, hybrid retrieval fusion with a committed corpus fixture, and a PII detector adapter with live-call-path coverage.

2026-08-09 — 5dacb73c… — first reading. Screened before reading; the tree was read, never installed, and no test was run.