Integrate accepted bus, backend, CLI, phase and birth-marker deletions

.github/workflows/debt-ratchet.yml

@@ -138,7 +138,10 @@ jobs:
           python-version-file: pyproject.toml
 
       - name: Install agent-comms
-        run: python -m pip install -e '.[dev]'
+        run: |
+          python -m pip install uv==0.12.0
+          uv sync --locked --all-extras
+          echo "$GITHUB_WORKSPACE/.venv/bin" >> "$GITHUB_PATH"
 
       - name: No new structural debt against the base
         env:

plans/REMAINING-SLOP-PLAN.md

@@ -133,13 +133,87 @@ Three audits, one per catalog rung group, each finding read in the code and chec
 - `RecoveryCondition` matching a lifecycle by `declared_name` string equality across two families (`recovery_states.py:21-26`): `issubclass`.
 
 **6.4 Rosters and registries beside families (MEMB-1/2/4, about 120).**
-- Five hand-rolled `__init_subclass__` registries: `Carry.declared` (`installer/carry.py:45-54`), `ThreadRename.kinds` (`thread_management.py:557-567`, gone with 4.1), `BirthlessIncarnation.members` (`thread_identity.py:185-192`), `CliToolArguments.types` (`tool_calls.py:89-95`), `_start_phases`/`_end_events` (`pi_events.py:59-93`): `DeclaredFamily` or attributes through the MRO.
-- `PI_0_85_1_EVENT_KINDS`, 36 names re-typed beside `NativePiEvent.known_kinds()` (`pi_native_backend.py:63-81`, check at `agent_backend.py:428`): derive the union; the equality check goes.
-- `THREAD_CAPABILITIES` plus `serves_threads`/`serves_new_threads` flags plus `validate_backends` probing with `getattr` (`agent_backend.py:358-466`): `ThreadServingBackend(AgentBackend, Steers, Compacts)` and `NewThreadBackend`; ABC instantiation enforces the roles.
+
+Backend capability slice (#2623, integration base `acb2ddd0e`, 2026-10-23):
+deleted `THREAD_CAPABILITIES`, both serving flags, `validate_backends` and the
+old default-selection method. `ThreadServingBackend` inherits `Steers` and
+`Compacts`; both record default factories invoke `NewThreadBackend.selected`.
+Capability bases precede `AgentBackend`, whose concrete refusal method would
+otherwise hide the abstract turn obligation. Sealing attaches only concrete
+locations after validating unique ownership. An incomplete declaration cannot
+rebind a location. `NoSessionFile` constructs the shared `UnstartedRuntime`;
+both copied backend `unstarted_type` declarations are deleted.
+NRA census covers 1,528 Core/hooks/tools/tests and paired Toad modules:
+6,973 original classes, 6,752 projected classes, zero parse omissions; the
+221 unprojected originals remain visible. Source evidence is not a behavior
+proof. Afterward: 1,529 modules, 6,974 original/6,753 projected classes, zero
+parse omissions; Core remains 953 modules/5,341 original classes. The extra
+dependency module/class came from the concurrently edited Toad source tree.
+The before/after receipts remain in the owned backend-capability-rosters
+scratch directory. Focused source checks pass 27 cases (25.69s), including
+fresh CLI/UI imports, typed native/durable locations, ABC refusal without
+location mutation, worker backend reuse and a real durable host reading 260
+saved rows while its SQLite writer stays open. The initial 24-pass/3-failure
+run exposed an untyped session-file fixture, a stale caller roster and missing
+local Node dependencies; all three were corrected, without counting that run
+as green. Installed copied-agent DM/reply/tab-return qualification belongs to
+the later paired integration; this branch has not changed the live runtime.
+Review corrected the declaration tests: they now re-execute the actual backend
+declaration module through standard `runpy.run_path` in a fresh namespace,
+declare real nominal subclasses and invoke the actual seal lifecycle. They do
+not monkeypatch production sealing, registry contents or backend locations.
+The abstract-class exclusion/no-rebind, competing-location and zero/two-default
+cases pass (3 cases, 0.11s); fresh-entrypoint import checks also pass. Production
+is unchanged from the 27-case qualification; the class census above describes
+that implementation checkpoint before the test-fixture correction.
+The independent Pi producer/converter event comparison is retained;
+deriving expected producer events from the converters would erase its check.
+This slice does not close `RostersPhase` or edit its evidence declaration.
+
+- Four remaining hand-rolled `__init_subclass__` registration owners: `Carry.declared` (`installer/carry.py:42-58`), `ThreadRename.kinds` (`thread_management.py`; still present, not removed by 4.1 at this integration base), `BirthlessIncarnation.members` (`thread_identity.py:214-226`), and `_start_phases`/`_end_events` (`pi_events.py:59-93`). Move registration through the existing nominal mechanisms or derive behavior through the MRO; preserve each boundary's actual keys and consumers. `CliToolArguments.types` and its private hook are removed by the CLI argument slice below.
+- Preserve the independent Pi producer ABI contract: `PI_0_85_1_EVENT_KINDS` in `pi_native_backend.py` declares the pinned runtime's emitted kinds, while `NativePiEvent.known_kinds()` declares converter coverage. `AgentBackend.require_event_family` compares these independent answers and reports missing or extra converters. Deriving the expected producer kinds from converters would erase this check. Refactor the internal `_start_phases`/`_end_events` registration lifecycle without weakening producer/converter coverage; removal of the external manifest is not a completion condition.
+- Backend roles: removed by #2623 through `ThreadServingBackend(Steers, Compacts, AgentBackend)` and declaration-derived `NewThreadBackend.selected`; abstract methods enforce thread obligations. `PiDurableBackend` remains a non-thread backend.
 - Display names re-typed from class names in about 40 places (`tools.py` labels, `notification_progress.py:45-157`, `todos.py:1502/2012/2088/2605/2675`, `display_order.py:76-136`, `relationships.py:382/397`, `turn_phase.py:169/218`, `config_options.py:192-194`, `input_attempt.py:400/440/447` `public_status`): derived from the declared name, literal only where the text differs.
 - Sigils `#`/`@` spelled by hand at about 22 sites (`catalog_document.py:61-232`, `channels.py`, `channel_targets.py`, `goal_presentation.py:71-72`, `goal_actions.py:305/318`, …): owned by `Tag.channel` and `ThreadTarget.mention`.
 - Store and file names restated outside their class: `root / "wire"` in about 20 places across 17 modules (the virtual bus plan's done-grep for `_wire_lock_path` misses every `_store_lock(root / "wire")`; its grep gains that pattern), `~/.local/state/agent-comms` six times, bridge inbox dirs, `"registry.json"` at `installer/upgrade.py:566`, `".registry-owner-guard"` at `goal_history.py:208`, `active-route.json` five times, Pi `auth.json` in four modules: `X.path_in(root)` and one state-home owner.
 
+CLI argument slice (#2626, base `acb2ddd0e`, 2026-10-23): the existing
+`RegisteredWireMeta` now owns argument-model registration after Pydantic/ABC
+construction. Each concrete declaration supplies its genuine external
+`cli_names` tuple as the registry key; decode derives matching shapes from
+that registry without retaining another alias map. Both production consumers,
+Claude `ToolUse.call` and OpenCode `ToolState.call`, keep the shared decode
+entrypoint. The seven shapes retain all 13 case-sensitive external spellings,
+extra input, canonical previews and edit patches. Unknown names and invalid
+arguments remain the original call; validation refusal belongs to its runtime.
+Abstract specializations cannot overwrite concrete aliases. The existing
+registry rejects duplicate tuples at declaration; overlapping nonidentical
+tuples raise an explicit ambiguity at decode, before argument validation can
+turn the conflict into an ordinary refused call. Overlap is not checked at
+declaration. A new shape still needs only its class and alias declaration;
+the improvement is shared declaration lifecycle and collision behavior, not
+a claimed reduction from multiple authored rosters.
+
+NRA before/after source receipts cover Core/hooks/tests/tools, paired Toad and
+the installed metaclass-registry/Pydantic source: 1,643/1,644 modules,
+7,434/7,438 original classes, 7,144 projected classes in each, zero parse
+omissions. The 290/294 unprojected originals remain explicit; the four added
+classes are function-local declaration controls. Dynamic source projection
+does not prove their runtime behavior. The compact index also leaves the
+`AutoRegisterMeta` package re-export unresolved as a base of
+`RegisteredWireMeta`; source inspection and a fresh-process runtime MRO read
+confirm its composition with `ModelMetaclass`. Focused source checks pass 31 cases
+(0.63s), exercising real Claude/OpenCode models, edit presentation and actual
+fresh declaration lifecycles without patching registry state. The first run
+passed 26 and failed two new assertions that incorrectly put call identity on
+search preview records; identity is checked on its owning `ToolCall` and
+emitted transcript instead. Receipts remain in the owned
+`cli-tool-argument-family` scratch directory. Fresh CLI/UI-model import checks
+also pass (two cases, 3.73s). This slice is a published draft,
+not merged, installed or live-qualified. Shared phase proof belongs to its bus
+owner; the other registries and the remaining name/sigil/path work keep 6.4
+open.
+
 ### Wave 7: records where parameter lists and dicts stand in
 
 - `ManagedLaunchRequest`: nine parameters declared in seven signatures and forwarded six times (`native_child.py:222/341`, `native_pi.py:491/536`, `native_session_prepare.py:67/88`, `backend.py:112`, `stream_agent_events` has 23), with one lock-launch-fence procedure written twice. About 60.
@@ -226,7 +300,7 @@ Waves 4 to 8 were written on 2026-10-21 and declared no phase in `plan_phases.py
 | wire_renames | 4.6 | no `WireName("admission…` in src | one unit, Toad in lockstep |
 | owned_dispatch | 5.4 | no `TurnInboxReceiver`, no hand `separators=(",", ":")` | one unit |
 | kind_families | 6.3 | no `TagAction`, `ExecutionOrigin`, `WhenBusy`, `ControlClassification`, `MembershipChange` or `ViewKind` enum classes | by file |
-| rosters | 6.4 | no `PI_0_85_1_EVENT_KINDS`, `THREAD_CAPABILITIES` or `serves_new_threads` | one unit |
+| rosters | 6.4 | internal registries and capability/name/path rosters derive from their owners; retain independent Pi producer/converter coverage | by family; shared phase proof has one owner |
 | records | 7 | no `ConnectivityFacet`/`observe_connectivity`; no `nonlocal` in the five document updaters | one unit |
 
 9.4 is Tristan's ruleset change. 9.5, 9.6, 9.9, 9.10 and 9.11 are claude-ai's and are measured by their own numbers in this section: a touched-tests PR check, the full ratchet list in the job summary, commander calls under 3 s, one GitHub reader, and one stack lock. The order across plans stays Tristan's: the virtual bus batch first, then multiple inheritance, the commander-native plan, then this plan. Phases with no stored data run in parallel with all of them on disjoint files.

plans/VIRTUAL-BUS-PLAN.md

@@ -88,10 +88,41 @@ Everything lands as one batch through the merge queue, drop-on-fail. Then one in
 
 Each phase, its predecessors and the evidence on `origin/main` that proves it done are declared in `src/agent_comms/plan_phases.py`; the commander's plan-idle rule and order refusals read that family, and accepting a task never marks a phase done.
 
-- `git grep -n "offset\|max_bytes\|bus.jsonl\|bus_meta\|_wire_lock_path\|has_private_wire_fields\|/ \"wire\"" src` returns no bus hits (the last pattern catches the hand-spelled `root / "wire"` lock path at about 20 sites).
+- `BusDeletionPhase.evidence` finds the five deleted modules absent and no known physical bus symbols, filename literals or lock-path expressions in Python source. Its AST check reads executable uses, including `/ "wire"`, `.with_name("wire")` and `.joinpath("wire")`, without matching prose or its own pattern declaration. This is a known-syntax screen; aliases and dynamically constructed paths still require source review. Remaining byte offsets and budgets are judged by their actual owner, since external native transcripts stay files.
 - The deleted modules are gone.
 - A process killed mid-publish leaves the whole message or none of it.
-- A reader that meets a held bus lock waits for it and never ends a turn (step C). Today `NativeSourceCursor._advance_proven` (`native_source_cursor.py:83`) takes the boundary with `blocking=False`, and with no contention allowance when no input is committed (`:79-80`). So any concurrent writer ends the turn with `BlockingIOError`: three times on core-5d8da75aa between 12:20 and 12:39 on 2026-10-21, with no install running.
+- Concurrent bus publication must not abort native cursor settlement. The current path is `CoordinatedTurn.cursor_status_for` → `CursorDelivery.advance` → `CursorOwner.advance_async` → `CursorOwner._advance_owned`. Bus reads use `Bus.observing` and SQLite snapshots; the cursor owner commits through the coordinator session's existing write transaction. `Coordination.run_async` runs this work through its worker with the default `StoreWait` contention policy. Qualify that path under competing writers within its bounded wait; no physical bus lock is reintroduced.
 - Every history converts with every record round-tripping.
 - On every machine: Toad pages history, delivery wakes units, claims hold, and a real message to a live agent is answered.
 - Lines deleted are reported first; the census confirms.
+
+Historical regression evidence: on core-5d8da75aa, the former
+`NativeSourceCursor._advance_proven` in `native_source_cursor.py` took its
+boundary with `blocking=False`, without contention allowance when no input
+was committed. Concurrent writers ended turns with `BlockingIOError` three
+times between 12:20 and 12:39 on 2026-10-21, with no install running. That
+class and file have since been deleted; the incident motivates the current
+contention obligation above, rather than describing a current source site.
+
+## Deletion checkpoint from the current integration branch
+
+At `acb2ddd0e`, all five old modules are already absent. Three physical bus locks
+remain: `NativeReplyPreparation.finished`, `SendRefusal.record`, and
+`CollaborationLedger.merge`. The last uses `.with_name("wire")`, which the old
+grep misses. The old grep also matches its own `bus_meta` pattern declaration,
+so its result does not establish whether physical bus uses remain.
+
+The deletion branch removes those locks through the existing owners. Unread
+publication keeps its local preparation lock, source revision check and SQLite
+write transaction; a busy SQLite writer leaves completed readers pending.
+Refusal recording keeps maintenance permission under the registry lock until
+its SQL transaction finishes. Ledger merge keeps the registered-author snapshot
+through the existing document update, in the declared registry-before-ledger
+lock order. None creates a replacement bus lock.
+
+The phase check reads committed source through `MainTree`. Its controls include
+the actual remaining path spellings, physical filenames and removed symbols,
+alongside comments and docstrings that describe deletion. This source checkpoint
+does not mark the phase complete on main or qualify an installed runtime. The
+merger still checks the affected copied native/Toad journey, including a fresh
+DM, its actual reply and tab return.

pyproject.toml

@@ -129,3 +129,6 @@ extend-immutable-calls = ["agent_comms.store_files.ExclusiveLock"]
 warn_return_any = true
 warn_unused_configs = true
 disallow_untyped_defs = true
+
+[tool.uv.sources]
+metaclass-registry = { git = "https://github.com/OpenHCSDev/metaclass-registry", branch = "main" }