Tasks: one owner per work item, one session per working tree

The last spec artifact. A DAG derived from the design and testing plan.

Task list

• [x] 1. Write the failing tests for the ownership rule

  • A same-repository pull request event delivers into the work item's session, spawns nothing, and is still recorded on the record with an empty tmuxTarget.
  • A same-repository pull request that already has an endpoint session stops being fed, with nothing torn down.
  • Depends on: none
  • Requirements: R1.1, R1.2, R1.3
  • Test: T1, T10 — pytest cli/tests/test_routing.py (red→green)

• [x] 2. Write the failing tests for "a checkout, or no session"

  • A cross-repository pull request with no workspace is declined, its event delivered into the work item's session, and its binding still recorded.
  • A cross-repository pull request with a workspace spawns in a worktree of that repository, keyed on its own slug and at a different path from the record's.
  • Depends on: none
  • Requirements: R2.1, R2.2, R2.4
  • Test: T8 — pytest cli/tests/test_routing.py -k cross_repo, T4 — pytest cli/tests/test_workspace.py (red→green)

• [x] 3. Capture the red run as evidence

  • Depends on: 1, 2
  • Requirements: R3.1
  • Test: evidence/red.md

• [x] 4. _same_repository, and the rule in _endpoint_for

  • One module-level helper comparing (provider, path); the rule placed beforerecord.endpoint_for(pr) so an existing endpoint session is bypassed too.
  • Depends on: 3
  • Requirements: R1.1, R1.2
  • Test: T1, T10 (green)

• [x] 5. _endpoint_cwd, and the spawn seam that uses it

  • Optional[str]; None is the refusal. Workspace-prepared checkout keyed on the pull request's slug; WorkspaceError caught and turned into a refusal rather than a raise. endpoint.cwd records the checkout actually used; graphlink.on_pr_spawn keeps record.cwd (the spec chain's home, not the session's).
  • Depends on: 3
  • Requirements: R2.1, R2.2, R2.3
  • Test: T4, T8 (green)

• [x] 5a. Self-review finding: enforce the invariant, do not infer it

  • _prepare_workspace falls back to spawnWorkdir for a payload naming no repository, which can resolve to the record's own tree — so the design's claim ("two harness conversations never share a working tree") held by inference from the workspace layout rather than by a check. _same_path + a shared-worktree refusal now enforce it, with a test driving the guard through the real spawn seam.
  • Depends on: 5
  • Requirements: R2.2, R2.4
  • Test: T8 — pytest cli/tests/test_routing.py -k lands_on_the_records_tree (red→green)

• [x] 6. session.pr_session_declined in the event catalogue

  • Named reasons no-separate-checkout / workspace-failed; session.pr_spawned's description amended to say it is now reached only from another repository.
  • Depends on: 5
  • Requirements: R2.4
  • Test: T12 — pytest cli/tests/test_eventlog.py (the catalogue-drift check)

• [x] 7. Rewrite the two integration scenarios the behaviour change invalidates

  • test_pr_comment_reaches_the_linked_issues_work_item and test_pr_event_still_reaches_its_work_item_after_the_link_is_removed, Gherkin included — the second still proves issue-172's durable binding, now by showing the event reaches the work item off the binding alone.
  • Depends on: 4, 5
  • Requirements: R1.1, R1.3, R1.4
  • Test: T2 (green)

• [x] 8. Run the full suite and the repo's own gates

  • Depends on: 4, 5, 6, 7
  • Requirements: R3.1
  • Test: T12, T13

• [x] 9. Capability docs and the config reference

  • docs/capabilities/webhook-triggers.md (the sessionPerPr clause and a changelog row), docs/config/cli/routing-options.md (tmux.sessionPerPr), docs/capabilities/capabilities.md if the index states the old shape.
  • Depends on: 8
  • Requirements: R1, R2
  • Test: T12 — pytest cli/tests/test_docs_parity.py

• [x] 10. Decision record and CHANGELOG

  • decision-088 refining decision-064 D3, indexed in decisions.md.
  • Depends on: 9
  • Requirements: R1, R2
  • Test: T13