[how2how2how2-arch]

Adds the step-by-step tutorial #26 asks for: docs/tutorials/langgraph-triage.md, plus a test that runs every command in it.

What the page does. One small LangGraph triage agent, built in the order the issue names rather than as one lump:

1. write the agent (scripted chat model, one instrumented tool, fully offline),
2. record a run — python triage_agent.py → a .sftrace bundle,
3. validate it — stepfork validate triage.sftrace --verify-integrity, then inspect to show what was captured,
4. replay it — stepfork replay triage.sftrace --entrypoint triage_agent:run_agent, every recorded boundary substituted,
5. write the expectation by hand, export a test against it, watch the buggy entrypoint fail with behavior mismatch at 'decision': expected "escalate", got "queue",
6. correct the decision, export against the fixed entrypoint, watch it pass.

It links to docs/langgraph.md for the adapter semantics and to examples/langgraph_agent for the two-tool example that also diffs and verifies integrity, and restates neither. The scenario is a support-ticket triage agent, deliberately not the refund agent the example already ships. mkdocs.yml gains one nav line for it.

"The documented commands are tested." tests/integration/test_docs_langgraph_tutorial.py extracts both code blocks from the page (the agent, and the correction introduced in step 7), concatenates them, and runs every documented command in a clean subprocess — record, validate with --verify-integrity, inspect, replay, export, and pytest on both the buggy and the fixed entrypoint — asserting the recorded event types, the substituted boundaries, the MATCHED verdict, and the two opposite pytest outcomes. A renamed flag or a removed option fails this test instead of a reader. It is modelled on the existing tests/integration/test_docs_langgraph_example.py, which guards the guide the same way.

What I verified, and what I could not. The full suite passes: 703 passed with this test, 701 passed with it excluded, so nothing else moved. Every output quoted in the page is copied from an actual run on this machine; the numbers that vary per run (event count, timings) appear only inside blocks that are labelled as output of a reference run, and the test asserts the behavioural facts, not those counts.

Not run locally: ruff check, ruff format --check, mypy, and mkdocs build --strict. This machine's wheel CDN throttles to about 19 KB/s, so uv sync could not fetch those wheels at all — I ran the suite out of a venv built from the cached runtime dependencies instead. Line lengths are within the 88-character limit and imports follow the configured isort order, checked by hand, but that is not ruff and I am not claiming it is. The three relative links (../langgraph.md, ../pytest-export.md, ../trace-format.md) resolve to files that exist. CI will be the first place those four checks actually run, and if one of them objects I will fix it rather than argue.

Refs #26.