ci(integration-test): promote qp-topic into the required suite

[modelmirror]

Maintainer-gated surfaces (promotion gate class) — ready for maintainer review and merge; not self-merged by design.
This PR touches scripts/promotion-gate.sh and .github/workflows/, both of which AGENTS.md holds for the maintainer even into staging.

Artifact served

The promotion gate's coverage of the qp-topic labeling path. qp-topic exercises the shared qp-topic-measure composite — the post-label half of the paid labeling run: the no-output guard, the fedcourts qp-topics publication gate, and the publish-and-validate path. Until now that path could regress and still promote. The scenario spends no tokens and binds no credentials, so requiring it costs a promotion nothing.

The scenario has a string of green live runs behind it — a standalone qp-topic run and several whole-suite scenario=all runs — which is the evidence its introduction deliberately deferred the gate promotion for.

What changed

Four synchronized sites, each pinned by an existing workflow-shape test (tests/test_workflow_promote.py fails on any partial update):

1. scripts/promotion-gate.sh — qp-topic added to REQUIRED_SCENARIOS (8 → 9 entries).
2. .github/workflows/integration-test.yml — {"scenario": "qp-topic", "engine": "claude-code"} added to the all matrix (7 → 8 legs; collect still rides the run as its own environment-free job, not a leg). The scenario dispatch input already listed qp-topic.
3. .github/workflows/promote.yml — qp-topic added to the per-scenario dispatch commands the freshness failure prints.
4. docs/testing.md, docs/pipeline.md — the "outside the required suite" qualifiers dropped and the counts corrected.

New true counts: 9 required runs = 8 all matrix legs + the collect job, with engine-smoke counted once per each of three engines.

Freshness matches the new leg without further change: for an engine-less entry the gate builds ^integration-test: qp-topic / and requires the @ staging$ suffix, which the leg's title integration-test: qp-topic / claude-code @ staging satisfies.

Pre-existing drift retired in the same paragraphs

• docs/testing.md still said "The collect scenario is outside the gate" and gave the required-run count as "seven … five of the six real scenarios". Both were already stale against the pre-diff 8-entry REQUIRED_SCENARIOS — collect joined the suite earlier. The sentence is dropped (its still-true half, that collect binds no environment and so no deployment-branch policy can refuse it, already lives in the scenario inventory above it) and the count rewritten.
• The qp-topic leg's own rationale comment in integration-test.yml argued from its non-required status — a premise this commit deletes. The justification inverts rather than merely losing a word: the environment binding the leg never reads is exactly what makes its run title's @ staging suffix a deployment-gated provenance claim, which is how freshness reads every required matrix leg. A collect-style environment-free job would have weakened the evidence.

Checks run

Python surface untouched (docs, workflows, one shell line), so the gate stages were scoped accordingly:

• uv run pytest tests/test_workflow_promote.py tests/test_workflow_collect.py tests/test_workflow_cell_invariants.py — 50 passed
• uvx zizmor@1.26.1 --persona=regular .github/workflows .github/actions — no findings
• pinned actionlint 1.7.12 with shellcheck -e SC2016 — clean
• shellcheck -e SC2016 scripts/promotion-gate.sh — clean

Review

workflow-reviewer and docs-reviewer both ran on the diff. No blockers from either. Both raised the same recommended finding — the now-contradictory qp-topic rationale comment — which is fixed above using workflow-reviewer's stronger framing. docs-reviewer's nits (the "every scenario" phrasing, which would have read as including the all option, and ragged reflow at the edit seams) and workflow-reviewer's nit on the environment parenthetical in docs/pipeline.md are all applied. workflow-reviewer independently confirmed against gh run list that a qp-topic / claude-code @ staging run has concluded success from the staging branch, and that the last green all predating this change cannot leak into the gate — freshness pins on head_sha, so a run at the promoted head executes the 8-leg matrix at that head.

Closes #1210 — note this does not fire from a staging merge; the orchestrator closes the issue by hand once this lands.