[cam1se]

Summary

Stacked on #39 (single-PR and no-PR mode, stack #46): merge #39 first. This PR's live tests expect the one-row single-mode view that #39 introduces.

Adds an end-to-end suite that runs the real gh and gh stack against live fixture PRs (issue #23), and fixes three classifier bugs the real binaries exposed:

• gh 2.101.0 installs a missing stack extension on first use and exits 0 without running the command. The adapter now retries once after seeing the install message; before, the action silently did nothing.
• gh stack exit 2 also means "not a git repository". Push/sync now map to LocalTrackingMissing only when stderr says the branch is not part of a stack, so a non-repo failure no longer offers to link.
• gh exit 4 (no token) now maps to Auth; the old text match missed gh's wording.

The fixture lives in this repo under fixture/* (stacks s1 three-PR, s2 conflict, s3 merged, s4 unstacked, plus fork PRs sharing a head branch name), with fixture-start/* reset tags, single-mode fixtures (s4 single PR on main, s5 single PR on a non-default base, s6 no-PR branch, s7 closed, s8 merged), scripts/live-fixture.sh reset|conflict|status, and rulesets that block deleting fixture branches and updating the tags. CI and commit-lint now skip fixture/* pull requests. make live-test runs the suite; a nightly (06:17 UTC) and dispatch-only workflow runs it in CI against a pinned known-good gh/gh-stack pair and the latest releases, so a latest-only failure signals upstream drift. A failed nightly run opens or updates an issue.

Testing

• cargo fmt --all -- --check: clean (cargo fmt --all run).

• cargo clippy --workspace --all-targets --locked -- -D warnings: clean.

• cargo test --workspace --locked --no-fail-fast: all pass (live tests skip without STACK_STAT_LIVE=1).

• Live: STACK_STAT_LIVE=1 cargo test -p stack-stat-core --test live_gh against this repo: 15 passed in ~150-225s, after fixing two failures the run found (exit-4 auth, and a fork test that did not prove owner disambiguation).

• TUI smoke via make demo: N/A because dashboard rendering is unchanged; the change is at the gh adapter boundary, which the live suite exercises directly.

• scripts/live-fixture.sh check passes; rebuild verified by deleting fixture/s6-nopr and rebuilding it. The PR-creation paths of rebuild (stacks s1/s2, PR s5) are untested.

Docs and architecture

• Living docs: new 07 Live validation; 05, docs index, and AGENTS.md updated.
• ADR-0004 (Proposed). The no-PTY rule in AGENTS.md is kept.
• cargo doc --workspace --no-deps --locked: builds.

Notes for reviewers

• Workflow verified in CI. A temporary branch push trigger (since removed) ran both matrix jobs green with the existing GH_ACCESS_TOKEN secret: pinned gh 2.101.0 + gh-stack v0.1.1 and latest gh 2.102.0 + gh-stack v0.2.1, 12/12 live tests each. The nightly schedule and manual dispatch start working once this merges; the failure-issue step only fires on scheduled failures and has not been exercised.
• Fixture PRs stay open permanently and carry the fixture and do not merge labels (fixture: a #27-fixture: c #29, fixture: s2 a #31, fixture: s2 b #32, fixture: s4 upstream PR (same branch name as fork PR) #37, fixture: fork PR sharing head branch name with stack s1 top #40, fixture: s5-single (single PR, non-default base) #43 open; fixture: s3 a #34, fixture: s3 b #35, fixture: s8-merged (single PR, non-default base) #45 merged and fixture: s4 fork PR (same branch name as upstream PR) #38, fixture: s7-closed (single PR, non-default base) #44 closed as permanent end states). Closing and reopening them would send notification emails, and the local TUI tests (Agent-led TUI tests of the real dashboard against the live fixture #42) need them open. Muting the threads needs the notifications token scope, which I did not have; mute them from account notification settings.
• Fixture drift. s1 is based on main, so gh stack sync can rebase it when main moves; tests reset before and after.
• Gaps (listed in the architecture page): discovery over 100 entries, fork clone pushing to its own fork, exact conflict exit code, real dashboard rendering, older gh versions.
• Existing fake-based tests are untouched except the classifier tests, rewritten to the observed output.
• Live suite is slow (~3 min) because mutating tests reset the fixture twice.

Closes: #23