fix(gh): classify pr checks by status column, not name substring

[john-chen-opus]

Problem

format_pr_checks (the rtk gh pr checks summarizer) classifies each row with a whole-line substring test:

if line.contains("[ok]") || line.contains("pass") { passed += 1 }
else if line.contains("[x]") || line.contains("fail") { failed += 1; … }
else if line.contains('*') || line.contains("pending") { pending += 1 }

gh pr checks prints tab-separated rows — <name>\t<bucket>\t<elapsed>\t<url> — so matching the whole line has two correctness bugs:

1. A check is classified by its NAME, not its status. A check named compass-audit or fail-fast smoke contains pass/fail as a substring and gets miscounted regardless of its real bucket.
2. skipping and cancel(ed) checks are dropped entirely. They match none of the three branches, so they vanish from the summary — the printed counts silently disagree with the real PR. (On a PR whose only non-passing checks are change-gated skips, gh still exits 0, so this summary is what the agent sees.)

This bit me while shipping an unrelated change: the summary disagreed with the actual check set, and the empty-checks case surfaced as a confusing "0 passed" rather than gh's real message.

Fix

Classify on the tab-separated status column, not a whole-line substring:

• Match the bucket via starts_with (absorbs skipping/skipped, cancel/canceled/cancelled).
• Fold skipped and cancelled into the totals (cancelled surfaces with failures — gh's own rollup exits non-zero for it).
• Lead with passed/total so the counts reconcile — skipped/pending checks are part of the total, never silently dropped.
• When no row parses (e.g. no checks reported on the '<branch>' branch), return empty so the caller falls back to gh's raw output instead of inventing a misleading 0 passed summary.

This aligns with the design philosophy in CONTRIBUTING.md:

• Correctness vs Token Savings — "when in doubt, preserve correctness";
• Transparency — the filtered output stays a faithful subset (counts match the real command);
• Never Block — the unparseable case falls back to raw rather than fabricating a summary.

Tests

Added three unit tests in gh_cmd.rs:

• test_format_pr_checks_classifies_by_bucket_not_name — names containing pass/fail are classified by their status column;
• test_format_pr_checks_reconciles_total_with_skips — 1 pass + 2 skips reports 1/3 with skips surfaced;
• test_format_pr_checks_empty_falls_back_to_raw — empty / "no checks reported" → empty (raw fallback).

Local: cargo fmt --check clean, cargo clippy --bin rtk clean (only a pre-existing rusqlite future-incompat note), cargo test --bin rtk cmds::git::gh_cmd → 69 passed.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}