Execute the serving cookbook's console transcripts, and mend their rot

[Shashankss1205]

What this does

docs/cookbook/06-serving-and-ops.md showed console transcripts that no test ever re-ran, and they had rotted exactly the way the issue predicted — and a little further:

• /healthz showed "version":"0.1.0" while grapharc/__init__.py says 0.1.1
• grapharc metrics has since grown events: and per_phase: fields (and nests per_node in --json) that the page did not show
• replay trace.jsonl <id> | tail -4 was a line short (pending = [])

This PR copies the marker pattern from tests/test_cookbook_models.py (<!-- verified: cli --> / <!-- verified: cli varies --> / <!-- needs-credentials -->) into tests/test_cookbook_serving.py, marks every console block on the page, corrects the stale transcripts, and re-records the serve transcript against the current tree.

Byte-compared vs varying, and why

Byte-compared (verified: cli) — deterministic output:

• the CLI tour (demo / trace / metrics / viz / replay / diff, 8 commands)
• the --json pair (metrics … --json and the exit 2 failure document)
• models openrouter/anthropic/claude-haiku-4.5 (pure spec resolution, no credentials)

These run through a real shell (bash -c, with a grapharc shim pinned to the test interpreter) as one session in one directory, because the later blocks read the trace file the tour wrote. Two spans the page itself declares variable are handled rather than skipped: the page's random run ids are remapped to the ids the re-run actually produced (in trace-file order), and wall-clock spans (duration_ms, (0.4ms) are masked. Every other character has to match — the demo block's leading ... elision compares the tail exactly.

Executed but not byte-compared (verified: cli varies) — genuinely machine-dependent:

• the serve transcript: binds a port and mints random session/run ids. Replayed end-to-end against a real server — the page's own mygraphs.py is written to disk, the page's own command starts the server on a free port, and every curl in the block must succeed against it (the deterministic banner and /healthz body are still compared, and the /healthz version line is additionally pinned to grapharc.__version__ by name so it cannot show 0.1.0 against an 0.1.1 tree again).
• models --check: probes this machine's credentials and PATH, and exits 1 when no real backend is usable, so only its shape is asserted — the same contract 02-models.md uses. This command was split out of the models <spec> block so the deterministic half stays byte-compared.

Never executed (needs-credentials):

• grapharc agent … — always needs a live model; the test asserts it shows no output it never printed.

An unmarked console block now fails the suite, so a future edit cannot put an unverified transcript on the page.

Gate proof

Changed tokens: 81 to tokens: 82 in the CLI tour's expected output: test_console_command_prints_exactly_what_the_page_shows[line1595-step3] fails with a diff. Reverted.

Verification

• uv run pytest tests/test_cookbook_serving.py — 57 passed
• uv run pytest — 1628 passed; the only failures are 5 pre-existing SIGALRM/max_seconds timing tests that fail identically on untouched main on this machine and are unrelated to this change
• uv run ruff check . — clean

Fixes #2

🤖 Generated with Claude Code