fix(pr-management-triage): suppress already-acted-on PRs with unchanged head SHA

[liwenjie200543]

Summary

• Implement the remaining acceptance criterion of Silent in-session de-dup of already-seen / already-skipped PRs #77: a PR already taken to a terminal action earlier in the same session is silently suppressed on re-encounter — no queue entry, no progress line, no prompt.
• The suppression is head-SHA aware, per the resolution proposed on the issue: an entry whose cached head_sha matches the freshly fetched head SHA has nothing new to read and is dropped; a changed head SHA does not match the suppression and falls through to the existing staleness rule (drop entry, re-classify). The two rules are complementary rather than conflicting.
• Extend the full-pagination loop pseudo-code, the Step 1 prose, the Session-cache invalidation section, and SKILL.md Step 1 accordingly; add three pagination-dedup regression cases (unchanged head suppressed / pushed head re-classified / cache entry without a terminal action_taken never suppresses) and restamp measured_tokens.

#1071 covered the pagination-level duplicate (a PR moving between pages); this patch covers the session-level one. The suppression set is session state and dies with the cache, so a fresh session re-surfaces everything.

Type of change

• Skill change (skills/<name>/) — eval fixtures updated below
• Tool / bridge contract (tools/<system>/*.md)
• Python package
• Documentation
• CI / dev loop

Test plan

• All 62 pr-management-triage eval fixtures assemble successfully in print mode (59 → 62: three new pagination-dedup cases)
• skill-eval framework tests: 155 passed; the 19 failing tests are subprocess-spawning cases that fail identically on unmodified main in this Windows environment (WinError 2 spawning POSIX-style commands) — verified against a clean main worktree, byte-identical failure list, so no regression is introduced by this change
• uv run --project tools/skill-token-count skill-token-count --write → measured_tokens: 5027
• tools/dev/skill-surface-hash.py --fix reports surface_hash in sync (no headings added, moved, or renamed)
• Spec-sync pre-check: tools/spec-loop/.last-sync (77884f5c) is a direct ancestor of the current tip, 16 commits behind; pr-management-family.md / triage-mode.md do not describe the pagination / session-cache layer, so no spec drift is introduced

Exact-automated (--cli) grading of the three new fixtures was not run locally (no model CLI in the authoring sandbox); the fixtures follow the existing pagination-dedup case shape and the expected outputs are pinned in expected.json.

RFC-AI-0004 compliance

• Vendor neutrality — no provider-specific behavior added

Linked issues

Closes #77 — implements the resolution proposed in this comment: suppress silently only while the head SHA is unchanged, re-classify when it changed.

Generative AI disclosure

ZCode (GLM) was used while preparing this change. The contributor reviewed the diff and test results before submission.

[Kaap10]

Thanks @liwenjie200543!

Went through the changes:

• Step 1 session suppression logic is clean and properly distinguishes unchanged head_sha from fresh pushes.
• Regression fixtures in pagination-dedup (cases 3–5) cover the edge cases nicely.
• Eval suite counts and token measurements are kept in sync, and all CI checks pass.

LGTM!