Tasks: wire the ingress to the process graph

Phase 3 of 3. Derived from design.md. DAG — a task may start once its dependencies are ticked. tdd.mode: standard: each task writes its failing test first.

flowchart TD
  T1[T1 Runtime.start] --> T4[T4 GraphLink class]
  T2[T2 spec_id_for] --> T4
  T3[T3 comments_from] --> T4
  T4 --> T5[T5 config block]
  T5 --> T6[T6 dispatcher call sites]
  T6 --> T7[T7 integration tests]
  T7 --> T8[T8 capability docs]

Task list

T1 — Runtime.start() — enter the start node

• Requirements: AC1, AC2, AC3
• Depends on: —
• Red: test_graph_runtime.py::test_start_enters_start_node_and_runs_entry_chain, ::test_start_is_idempotent_when_pointer_exists, ::test_start_persists_pointer_before_entry_chain
• Green: add start() to cli/the_loop/graph/runtime.py; add graph.started to eventlog.EVENT_TYPES (the drift test in test_eventlog.py enforces this).
• [x] Done

T2 — spec_id_for() — ref → spec-directory id

• Requirements: AC8, A5
• Depends on: —
• Red: test_graphlink.py::test_spec_id_for_github_ref, ::test_spec_id_for_other_provider_is_none
• Green: derive issue-<int> from the parsed ref; None for non-GitHub providers.
• [x] Done

T3 — comments_from() — payload → attributed comments

• Requirements: AC5, B1
• Depends on: —
• Red: test_graphlink.py::test_comments_from_each_event_shape, ::test_comment_without_author_is_dropped, ::test_unrelated_event_yields_no_comments
• Green: extract issue_comment / pull_request_review_comment / pull_request_review bodies with their authors; drop any entry missing an author or a body.
• [x] Done

T4 — GraphLink — the seam, with every skip path

• Requirements: AC4, AC9, AC11, AC12
• Depends on: T1, T2, T3
• Red: test_graphlink.py::test_disabled_link_does_nothing, ::test_missing_spec_dir_is_skipped, ::test_awaiting_start_is_skipped, ::test_runtime_exception_is_swallowed, ::test_on_event_advances_with_comments
• Green: cli/the_loop/graphlink.py — GraphLinkConfig, GraphLink.on_spawn, GraphLink.on_event; add graph.link_failed to eventlog.EVENT_TYPES.
• [x] Done

T5 — routing.graph config block

• Requirements: AC12
• Depends on: T4
• Red: test_graphlink.py::test_routing_config_parses_graph_block
• Green: GraphLinkConfig field on RoutingConfig.from_mapping; document the block in .the-loop/cli-config.schema.json and skills/the-loop/templates/cli-config.yaml.
• [x] Done

T6 — Dispatcher call sites

• Requirements: AC1, AC5, AC10, AC11
• Depends on: T5
• Red: test_graphlink.py::test_dispatcher_starts_graph_on_spawn, ::test_dispatcher_advances_graph_on_delivery
• Green: construct GraphLink in Dispatcher.__init__ (and rebuild it in the hot-reload path); call on_spawn after a successful spawn, on_event after a successful delivery.
• [x] Done

T7 — Integration tests (Gherkin)

• Requirements: AC2, AC6, AC10, AC11, A1
• Depends on: T6
• Red/Green: cli/tests/test_graphlink_integration.py — spawn-starts-graph, reviewer-approval-reaches-the-gate (authorized vs unauthorized), failing-hook-does-not- cost-the-delivery.
• [x] Done

T9 — Let a passing gate's verdict reach its edges

Discovered by T7, not planned. The first integration test to advance a real human-approval node parked with no edge from requirements-approval on 'pass'. ChainOutcome.outcome read the routing value only from a blocking result, but classify-feedback returns pass carrying data["outcome"] = "approved" — so the verdict was discarded and all three approval nodes in pdlc.yaml, whose edges are declared on: approved / on: changes-requested, could never route. Pre-existing since issue-109 and never caught, because every existing test calls the hook directly rather than through advance(). In scope: AC6 cannot hold without it.

• Requirements: AC6
• Depends on: T7
• Red: test_graph_chain.py::test_a_passing_hooks_explicit_outcome_is_what_edges_route_on, ::test_a_chain_of_plain_passes_still_routes_on_pass
• Green: ChainOutcome.outcome falls back to the last result that declared an explicit data["outcome"]; a plain HookResult.ok still reports pass.
• [x] Done

T10 — Refuse a checkout that is not the work item's own repo

Found by CI on this PR, not by design review. The gate reported three work items instead of one: running the suite had written graph-state.json into the real docs/specs/issue-1 and issue-15. Cause — the dispatcher tests use github:octo/repo#15, which maps to issue-15, and their session cwd is .. That is a production hazard, not test noise: with the default spawnWorkdir: "." an event about any repo's issue #15 drives the operator's own issue-15.

• Requirements: AC14, A6
• Depends on: T4
• Red: test_graphlink.py::test_a_checkout_of_another_repo_is_never_coupled, ::test_a_checkout_with_no_origin_is_skipped, ::test_a_directory_that_is_not_a_checkout_is_skipped, ::test_the_work_items_own_checkout_is_coupled
• Green: _checkout_belongs_to + _repo_slug in graphlink.py, failing closed; the test fixtures became real checkouts with an origin.
• [x] Done

T8 — Capability docs + spec fold-in

• Requirements: ready-to-ship gate
• Depends on: T7
• Green: update docs/capabilities/process-graph.md and docs/capabilities/cli.md with the coupling and its history row; keep execution-log.md current.
• [x] Done