Refresh SGLang review corpus to 2026 + backfill model PR histories

[BBuf]

What

A timeliness pass over the auto-driven skills, focused on the two areas that
were most stale and most verifiable.

1. SGLang humanize-review corpus → through 2026, full multi-round capture

• Extended coverage from project start → 2025-12-31 to project start →
  2026 (latest refresh). The corpus roughly doubled:

  metric	before	after
  threads	17,914	32,639
  comments	39,511	76,516
  PR years	2024–2025	2024–2026 (14,236 threads from 2026 PRs)
  max comments / thread	77	142

• Fixed a multi-round truncation bug. The collector fetched only the first
  100 GraphQL nodes of each PR's comments and reviews connections, silently
  dropping the tail of long discussions. Both now paginate fully via cursor
  follow-up queries — conversation_paginated_prs=8 and
  review_submission_paginated_prs=2 were previously truncated and are now
  captured end-to-end. This directly addresses the "comments / multi-round
  discussion capture may be incomplete" concern.

• Hardened the collector for long runs (a full scrape is ~40 min): it now
  reads the gh token fresh, refreshes it on 401/403 (OAuth tokens rotate
  mid-run), and retries transient connection resets (OSError/HTTPException)
  with capped backoff.

• Renamed the corpus to a date-agnostic sglang-review-corpus.jsonl.gz /
  .metadata.json and switched to a dynamic window label; updated SKILL.md,
  the repo README, the query-helper default, and the tests.

2. Model PR optimization history → dated backfill audits on all 62 docs

The canonical generator (tools/rebuild_model_pr_history_from_git.py) needs
worktrees + a warm cache + thousands of GitHub API calls to fully regenerate,
which would also rewrite every doc wholesale. Instead, each of the 62
model-family docs (sglang + vllm) gets a conservative, git-traced
## <date> PR Backfill Audit section prepended after the title. It lists
PR-numbered merges that touched that model's tracked implementation files
after the doc's previous freshness cutoff and are not yet folded into the
older timeline / diff-audit cards — 643 new PRs surfaced in total. PR titles
are normalized (agent-tag/emoji stripped) so the dossier-quality test stays
green. The convention is documented in the module README and SKILL.

Intentionally left as-is

The dated GPU/benchmark/profiler/architecture-diagram validation runs in the
other skills are point-in-time evidence (specific hardware + commits + metrics).
I did not invent fresh dates over them, since they can't be honestly "refreshed"
without re-running on hardware.

Validation

• pytest tests/ → 105 passed (incl. the humanize-review metadata test now
  asserting source_years == [2024, 2026], and the model-PR dossier-quality
  test across all backfilled docs).
• ruff / isort / black clean on the changed collector and query scripts.

🤖 Generated with Claude Code