fix(gateway): possibilities-matrix fast-path hardening + collision-context guards

The W7 fast path now rejects empty/trivial notes (its sole compensating
control for the skipped journal gates), pushes the branch before the
behind-base check, and pairs the local-gate fallback with the toolchain
guard; the WORK_ALREADY_DONE prompt no longer promises a fast path to
verifying tasks the gate routes elsewhere. build_collision_context now
degrades gracefully at all three call sites instead of breaking the
gate review, PM briefing, or collision-map route.

roboco/api/routes/tasks.py

@@ -1371,14 +1371,22 @@ async def get_task_collision_map(
         raise HTTPException(
             status_code=status.HTTP_404_NOT_FOUND, detail="Task not found"
         )
-    siblings = (
-        await service.get_subtasks(UUID(str(task.parent_task_id)))
-        if task.parent_task_id
-        else []
-    )
     # No actual files here — the panel shows the declared surface + sibling
     # overlap only; drift stays in the in-context evidence envelope.
-    ctx = build_collision_context(task=task, siblings=siblings)
+    # Best-effort: a fetch/build failure degrades to no siblings rather than
+    # a 500 — the route still returns the task's own declared surface.
+    ctx: list[dict[str, Any]] | None = None
+    try:
+        siblings = (
+            await service.get_subtasks(UUID(str(task.parent_task_id)))
+            if task.parent_task_id
+            else []
+        )
+        ctx = build_collision_context(task=task, siblings=siblings)
+    except Exception as exc:
+        _logger.warning(
+            "collision_map_route_skip", task_id=str(task.id), error=str(exc)
+        )
     return CollisionMapResponse(
         task_id=str(task.id),
         parent_task_id=str(task.parent_task_id) if task.parent_task_id else None,

roboco/runtime/orchestrator.py

@@ -14275,7 +14275,7 @@ If the fast path refuses (a gate it checks is not actually met), the
         # just collapses the 3-5-turn re-derivation to a single i_am_done call.
         if (
             settings.possibilities_matrix_enabled
-            and status in ("claimed", "in_progress", "verifying")
+            and status in ("claimed", "in_progress")
             and task.get("pr_created")
             and task.get("commits")
         ):

roboco/services/gateway/choreographer/_impl.py

@@ -1010,11 +1010,11 @@ class Choreographer:
             return None
         try:
             siblings = await self.task.get_subtasks(parent_id)
-        except Exception:  # best-effort: omit on any fetch failure
+            return build_collision_context(
+                task=t, siblings=siblings, actual_files=actual_files
+            )
+        except Exception:  # best-effort: omit on any fetch/build failure
             return None
-        return build_collision_context(
-            task=t, siblings=siblings, actual_files=actual_files
-        )
 
     async def _with_collision_briefing(
         self, briefing: dict[str, Any], full: bool, task: Any
@@ -2853,6 +2853,49 @@ class Choreographer:
             return None
         return await self._i_am_done_fast_path(ctx)
 
+    async def _fast_path_rejection(self, ctx: _IAmDoneContext) -> Any:
+        """The fast path's ordered rejection cascade; None means proceed.
+
+        Order matters: notes first (cheap, no writes), then finding
+        resolution + guards, then the open-findings re-check, then the
+        quality verdict with its toolchain backstop.
+        """
+        if soup := self._soup_reason(ctx.notes, "notes", 4):
+            return soup
+        await self._apply_resolved_findings(ctx)
+        guards = (
+            lambda: self._check_submit_qa_field_gates(
+                ctx.agent_id, ctx.task_id, ctx.task
+            ),
+            # Push-first, mirroring the standard gate: _behind_base_gate reads
+            # origin, which must already reflect the pushed head.
+            lambda: self._ensure_branch_pushed(ctx),
+            lambda: self._behind_base_gate(ctx),
+            lambda: self._conventions_gate(ctx),
+        )
+        for guard in guards:
+            if rejection := await guard():
+                return rejection
+        # FINDINGS_ADDRESSED re-checked post-resolution: a resolved_findings entry
+        # in THIS call may have closed the open set; anything still open blocks.
+        if await self._open_finding_ids(ctx.task_id):
+            return Envelope.invalid_state(
+                message="fast path blocked — open findings remain on the ledger",
+                remediate=(
+                    "name every open finding in resolved_findings "
+                    "({finding_id, note}), or leave possibilities_matrix off "
+                    "and run the standard i_am_done path"
+                ),
+                context_briefing=ctx.briefing,
+            )
+        rejection, ran_local = await self._fast_path_quality_verdict(ctx)
+        # The local gate ran (no CI signal) without a toolchain that can
+        # execute it is a hollow pass — pair the two like the standard gate
+        # does. Skipped when CI-green already decided (no local run at all).
+        if rejection is None and ran_local:
+            rejection = await self._toolchain_broken_guard(ctx.agent_id, ctx.task)
+        return rejection
+
     async def _i_am_done_fast_path(self, ctx: _IAmDoneContext) -> Envelope:
         """Work-already-done fast path: slimmed gates + direct transition chain.
 
@@ -2865,37 +2908,13 @@ class Choreographer:
         non-negotiable gates the predicate already asserts, plus conventions
         (leaf dev tasks skip ``awaiting_pr_review`` so the conventions check
         must not also be skipped here — it is tree-sitter, sub-second, 0 turns
-        when green).
+        when green). Since ``notes`` is the sole compensating control for the
+        skipped journal gates, it is enforced here directly (empty AND soup —
+        unlike the standard path's soup-only check, which skips an empty value
+        because presence there is gated by the journal requirements this path
+        doesn't run).
         """
-        await self._apply_resolved_findings(ctx)
+        if rejection := await self._fast_path_rejection(ctx):
-        guards = (
-            lambda: self._check_submit_qa_field_gates(
-                ctx.agent_id, ctx.task_id, ctx.task
-            ),
-            lambda: self._behind_base_gate(ctx),
-            lambda: self._ensure_branch_pushed(ctx),
-            lambda: self._conventions_gate(ctx),
-        )
-        for guard in guards:
-            if rejection := await guard():
-                return await self._reject_i_am_done(ctx, rejection)
-        # FINDINGS_ADDRESSED re-checked post-resolution: a resolved_findings entry
-        # in THIS call may have closed the open set; anything still open blocks.
-        if await self._open_finding_ids(ctx.task_id):
-            return await self._reject_i_am_done(
-                ctx,
-                Envelope.invalid_state(
-                    message="fast path blocked — open findings remain on the ledger",
-                    remediate=(
-                        "name every open finding in resolved_findings "
-                        "({finding_id, note}), or leave possibilities_matrix off "
-                        "and run the standard i_am_done path"
-                    ),
-                    context_briefing=ctx.briefing,
-                ),
-            )
-        rejection, _ran_local = await self._fast_path_quality_verdict(ctx)
-        if rejection is not None:
             return await self._reject_i_am_done(ctx, rejection)
         try:
             if str(ctx.task.status) == "claimed":

roboco/services/gateway/choreographer/pr_gate.py

@@ -1213,6 +1213,9 @@ class PRGateMixin(_Base):
             return None
         try:
             siblings = await self.task.get_subtasks(t.parent_task_id)
+            return build_collision_context(
+                task=t, siblings=siblings, actual_files=files_changed or None
+            )
         except Exception as exc:
             logger.warning(
                 "gate_review_collision_context_skip",
@@ -1220,9 +1223,6 @@ class PRGateMixin(_Base):
                 error=str(exc),
             )
             return None
-        return build_collision_context(
-            task=t, siblings=siblings, actual_files=files_changed or None
-        )
 
     async def _build_gate_review_evidence(self, t: Any) -> dict[str, Any]:
         """Inline evidence for claim_gate_review: the assembled diff +

tests/integration/test_task_collision_map_route.py

@@ -153,6 +153,42 @@ async def test_collision_map_shows_overlapping_sibling(collision_client: dict) -
     assert sib_entry["undeclared"] == []
 
 
+@pytest.mark.asyncio
+async def test_collision_map_degrades_on_builder_failure(
+    collision_client: dict, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    # build_collision_context is called outside the get_subtasks try/except at
+    # this site's call site; a raise there must still yield a 200 with no
+    # siblings, never a 500 — mirrors qa.py's claim_review evidence pattern.
+    client = collision_client["client"]
+    parent = _task(collision_client, title="parent", status=TaskStatus.PENDING)
+    await collision_client["db"].flush()
+    under = _task(
+        collision_client,
+        parent_task_id=parent.id,
+        title="under review",
+        intends_to_touch=["roboco/services/git.py"],
+    )
+    _task(
+        collision_client,
+        parent_task_id=parent.id,
+        title="colliding sibling",
+        intends_to_touch=["roboco/services/git.py"],
+    )
+    await collision_client["db"].flush()
+
+    def _boom(**_kw: Any) -> None:
+        raise RuntimeError("collision builder exploded")
+
+    monkeypatch.setattr("roboco.api.routes.tasks.build_collision_context", _boom)
+
+    response = await client.get(f"/api/tasks/{under.id}/collision-map", headers=_HDR)
+    assert response.status_code == HTTPStatus.OK
+    body = response.json()
+    assert body["siblings"] == []
+    assert body["intends_to_touch"] == ["roboco/services/git.py"]
+
+
 @pytest.mark.asyncio
 async def test_collision_map_omits_non_overlapping_sibling(
     collision_client: dict,

tests/unit/gateway/test_collision_context.py

@@ -4,13 +4,22 @@ Pins the truth table — no parent → None, no surfaced siblings → None,
 file-overlap sibling shown, both-migration shown without file overlap,
 shared-only-without-overlap NOT shown, declared-vs-actual drift computed,
 and the caps respected. Pure (no DB, no IO): duck-typed task/sibling rows.
+
+Also covers the two choreographer-level wrappers around it
+(``_gate_collision_evidence`` / ``_collision_context_for``): a raise from the
+builder itself — not just the sibling fetch — must still degrade to
+``None``/omitted, never break the caller (``claim_gate_review`` evidence /
+the planning briefing).
 """
 
 from __future__ import annotations
 
 from types import SimpleNamespace
+from unittest.mock import AsyncMock
 from uuid import uuid4
 
+import pytest
+from roboco.services.gateway.choreographer import Choreographer, ChoreographerDeps
 from roboco.services.gateway.choreographer.collision import (
     COLLISION_GLOB_CAP,
     COLLISION_SIBLING_CAP,
@@ -191,6 +200,80 @@ def test_sort_key_orders_by_priority_then_sequence() -> None:
     assert [e["sequence"] for e in ctx] == [5, 1, 0]
 
 
+def _make_choreographer(*, task_service: AsyncMock) -> Choreographer:
+    return Choreographer(
+        ChoreographerDeps(
+            task=task_service,
+            work_session=AsyncMock(),
+            git=AsyncMock(),
+            a2a=AsyncMock(),
+            journal=AsyncMock(),
+            audit=AsyncMock(),
+            evidence_repo=AsyncMock(),
+        )
+    )
+
+
+class TestGateCollisionEvidenceDegradesOnBuilderFailure:
+    """``_gate_collision_evidence`` (claim_gate_review evidence)."""
+
+    @pytest.mark.asyncio
+    async def test_builder_raise_omits_block(
+        self, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        t = _task()
+        task_service = AsyncMock()
+        task_service.get_subtasks.return_value = [_sib()]
+        c = _make_choreographer(task_service=task_service)
+        monkeypatch.setattr(
+            "roboco.services.gateway.choreographer.pr_gate.build_collision_context",
+            lambda **_kw: (_ for _ in ()).throw(RuntimeError("boom")),
+        )
+
+        result = await c._gate_collision_evidence(t, [])
+
+        assert result is None
+
+    @pytest.mark.asyncio
+    async def test_fetch_raise_still_omits_block(self) -> None:
+        t = _task()
+        task_service = AsyncMock()
+        task_service.get_subtasks.side_effect = RuntimeError("db down")
+        c = _make_choreographer(task_service=task_service)
+
+        result = await c._gate_collision_evidence(t, [])
+
+        assert result is None
+
+
+class TestCollisionContextForDegradesOnBuilderFailure:
+    """``_collision_context_for`` (the planning/claim briefing helper)."""
+
+    @pytest.mark.asyncio
+    async def test_builder_raise_returns_none(
+        self, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        t = _task()
+        task_service = AsyncMock()
+        task_service.get_subtasks.return_value = [_sib()]
+        c = _make_choreographer(task_service=task_service)
+        monkeypatch.setattr(
+            "roboco.services.gateway.choreographer._impl.build_collision_context",
+            lambda **_kw: (_ for _ in ()).throw(RuntimeError("boom")),
+        )
+
+        assert await c._collision_context_for(t) is None
+
+    @pytest.mark.asyncio
+    async def test_fetch_raise_returns_none(self) -> None:
+        t = _task()
+        task_service = AsyncMock()
+        task_service.get_subtasks.side_effect = RuntimeError("db down")
+        c = _make_choreographer(task_service=task_service)
+
+        assert await c._collision_context_for(t) is None
+
+
 if __name__ == "__main__":
     import pytest
 

tests/unit/gateway/test_possibilities_matrix.py

@@ -256,6 +256,7 @@ class _FastPathMocks:
     conventions: AsyncMock
     open_findings: AsyncMock
     record_milestone: AsyncMock
+    toolchain: AsyncMock
 
 
 def _stub_fast_path(
@@ -266,6 +267,7 @@ def _stub_fast_path(
     ok = AsyncMock(return_value="OK")
     reject = AsyncMock(return_value="REJECT")
     record_milestone = AsyncMock(return_value=None)
+    toolchain = AsyncMock(return_value=None)
     monkeypatch.setattr(c, "_apply_resolved_findings", AsyncMock(return_value=None))
     monkeypatch.setattr(c, "_check_submit_qa_field_gates", AsyncMock(return_value=None))
     monkeypatch.setattr(c, "_behind_base_gate", AsyncMock(return_value=None))
@@ -277,6 +279,7 @@ def _stub_fast_path(
         "_fast_path_quality_verdict",
         AsyncMock(return_value=(quality_rejection, False)),
     )
+    monkeypatch.setattr(c, "_toolchain_broken_guard", toolchain)
     monkeypatch.setattr(c, "_notify_qa", AsyncMock(return_value=None))
     monkeypatch.setattr(c, "_touch", AsyncMock(return_value=None))
     monkeypatch.setattr(c, "_record_milestone_progress", record_milestone)
@@ -288,6 +291,7 @@ def _stub_fast_path(
         conventions=conventions,
         open_findings=open_findings,
         record_milestone=record_milestone,
+        toolchain=toolchain,
     )
 
 
@@ -365,3 +369,113 @@ async def test_fast_path_ci_failure_rejects_before_transition(
     await c._i_am_done_fast_path(_ctx())
     stubs.reject.assert_awaited_once()
     c.task.submit_qa.assert_not_awaited()
+
+
+# ---------------------------------------------------------------------------
+# _i_am_done_fast_path: notes is the sole compensating control for the
+# skipped journal gates — empty AND soup are rejected (unlike the standard
+# path's soup-only check, which skips an empty value).
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_fast_path_empty_notes_rejected(monkeypatch: pytest.MonkeyPatch) -> None:
+    c = Choreographer(_deps())
+    stubs = _stub_fast_path(c, monkeypatch)
+    ctx = _ctx()
+    ctx.notes = ""
+    await c._i_am_done_fast_path(ctx)
+    stubs.reject.assert_awaited_once()
+    c.task.submit_verification.assert_not_awaited()
+    c.task.submit_qa.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_fast_path_trivial_notes_rejected(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    c = Choreographer(_deps())
+    stubs = _stub_fast_path(c, monkeypatch)
+    ctx = _ctx()
+    ctx.notes = "wip"
+    await c._i_am_done_fast_path(ctx)
+    stubs.reject.assert_awaited_once()
+    c.task.submit_verification.assert_not_awaited()
+    c.task.submit_qa.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_fast_path_real_notes_passes(monkeypatch: pytest.MonkeyPatch) -> None:
+    c = Choreographer(_deps())
+    stubs = _stub_fast_path(c, monkeypatch)
+    ctx = _ctx()
+    ctx.notes = "already implemented in a prior session"
+    await c._i_am_done_fast_path(ctx)
+    stubs.ok.assert_awaited_once()
+    c.task.submit_qa.assert_awaited_once()
+
+
+# ---------------------------------------------------------------------------
+# _i_am_done_fast_path: guard order — push before behind-base, mirroring the
+# standard gate (_behind_base_gate reads origin, which must already reflect
+# the pushed head).
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_fast_path_pushes_branch_before_behind_base_check(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    c = Choreographer(_deps())
+    _stub_fast_path(c, monkeypatch)
+    order: list[str] = []
+
+    async def _push(_ctx: Any) -> None:
+        order.append("push")
+
+    async def _behind(_ctx: Any) -> None:
+        order.append("behind")
+
+    monkeypatch.setattr(c, "_ensure_branch_pushed", _push)
+    monkeypatch.setattr(c, "_behind_base_gate", _behind)
+    await c._i_am_done_fast_path(_ctx())
+    assert order == ["push", "behind"]
+
+
+# ---------------------------------------------------------------------------
+# _i_am_done_fast_path: toolchain backstop pairs with the local quality gate
+# fallback (no CI signal) — never with the CI-green branch.
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_fast_path_toolchain_broken_blocks_on_local_fallback(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    c = Choreographer(_deps())
+    stubs = _stub_fast_path(c, monkeypatch)
+    monkeypatch.setattr(
+        c,
+        "_fast_path_quality_verdict",
+        AsyncMock(return_value=(None, True)),  # local gate ran (no CI signal), passed
+    )
+    stubs.toolchain.return_value = Envelope.invalid_state(
+        message="toolchain broken", remediate="fix"
+    )
+    await c._i_am_done_fast_path(_ctx())
+    stubs.reject.assert_awaited_once()
+    c.task.submit_qa.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_fast_path_toolchain_guard_skipped_on_ci_green(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    c = Choreographer(_deps())
+    stubs = _stub_fast_path(c, monkeypatch)
+    monkeypatch.setattr(
+        c, "_fast_path_quality_verdict", AsyncMock(return_value=(None, False))
+    )
+    await c._i_am_done_fast_path(_ctx())
+    stubs.ok.assert_awaited_once()
+    stubs.toolchain.assert_not_awaited()

tests/unit/runtime/test_possibilities_matrix_prompt.py

@@ -80,12 +80,16 @@ async def test_armed_claimed_with_pr_and_commits_is_work_already_done(
 
 
 @pytest.mark.asyncio
-async def test_armed_verifying_with_pr_and_commits_is_work_already_done(
+async def test_armed_verifying_with_pr_and_commits_is_not_work_already_done(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:
+    # verifying is excluded from the proxy's trigger set: _i_am_done_pre_gate_dispatch
+    # routes an owned `verifying` task to the resume path before the fast path is
+    # ever reachable, so steering the prompt there would be a dead-end promise.
     monkeypatch.setattr(settings, "possibilities_matrix_enabled", True)
     prompt = await _orch()._build_dev_prompt(_task(status="verifying"))
-    assert "WORK ALREADY DONE" in prompt
+    assert "WORK ALREADY DONE" not in prompt
+    assert "WORKFLOW STATE: VERIFYING" in prompt
 
 
 @pytest.mark.asyncio