feat(clp): speed up logtype-insights with a one-command bootstrap and model2vec clustering

[robinlee09201]

Summary

Cuts the logtype-insights skill's two dominant costs — kick-start latency and classification wall time — and makes every step visible to the user.

New helpers (shared bin/)

• logtype-insights-bootstrap — one command replaces the ~6 hand-composed pipelines of the old steps 3–5: schema sample, per-field value distributions (DIST lines), the stats.log_shapes dictionary dump, and the classification-cache probe, printed as a grep-able KEY=VALUE summary (LOGTYPE_COUNT, FALLBACK, CACHE_MODE, APP_KEY, TO_CLASSIFY, file paths). Handles the templatize fallback for pre-shapes-API binaries and converts the frequency table to canonical NDJSON so the cache probe works there too (previously it silently probed an empty file on 0.12.x). SIGPIPE-safe on high-distinct fields (a head -8 mid-pipeline used to kill the script with exit 141 under set -e + pipefail).
• logtype-cluster (setup / cluster / expand) — model2vec (potion-base-8M, numpy-only) pre-clustering. The classifier labels one representative per cluster by id; expand re-attaches member logtypes byte-exact from the cluster file, so GROWTH cache merging is safe by construction — the LLM never echoes logtype strings.

Skill changes

• skills-claude: the 621-line monolithic SKILL.md becomes a thin ~170-line workflow plus three on-demand references (logtype-baseline.md, logtype-classify.md, logtype-insight.md) following the plugin's progressive-disclosure pattern.
• skills-codex: stays a monolith per convention, trimmed to use the same two scripts with inline classification.
• Classification runs in one haiku-default subagent (sonnet retry on validation failure) instead of pasting every template into sonnet.
• Both variants keep the mandatory scoped semantic cross-check from Migrate clp skills to the clp-core 0.13 shapes API (stats.log_shapes) #5 and now narrate every step to the user, including all fallback/retry branches.
• Docs updated: helper lists (README, CONTRIBUTING), LOCAL_TESTING smoke block, release-testing walkthrough, dev preflight.

Test plan

• bash -n + shellcheck on the bootstrap; py_compile on the cluster script; claude plugin validate + scripts/validate-codex-plugin.sh pass
• Bootstrap on the vLLM demo archive: LOGTYPE_COUNT=934, FALLBACK=SHAPES_OK, cache probe modes NEW → UPTODATE verified
• Templatize fallback exercised on a 0.12.x binary (two-phase --message re-run), freqs → NDJSON feeds the cache probe
• Byte-exactness: expand output diff-identical to the input NDJSON; GROWTH merge on a grown archive returns UPTODATE on re-run
• SIGPIPE regression: bootstrap exits 0 on an archive with a ~20k-distinct timestamp field
• End-to-end claude --plugin-dir: id-only assignments, 8-section report