Generated SKILL.md files exceed 100k chars because ~20k of shared preamble is duplicated into every one #3199

What happens

Every generated SKILL.md in this repo opens with essentially the same ~20k chars of preamble
(upgrade flow, telemetry prompts, routing prompts, lake intro, artifacts sync) copied verbatim out
of SKILL.md.tmpl. Across the repo, 49.9% of all skill text is duplicated line-for-line between
skills, and seven SKILL.md files now exceed 100,000 characters:

(Median SKILL.md across all 58 files: 48,612 chars.)

Why it matters beyond file size

Several agent runtimes and skill loaders cap a single SKILL.md at 100,000 characters. One
concrete failure mode: the Hermes skill editor refuses every edit — patch, rewrite, anything —
on a skill whose SKILL.md exceeds 100k chars, and the rejection message does not mention size, so
the file silently becomes read-only. Independent of tooling, a 167k-char skill consumes a large
share of a context window when loaded on demand.

Measurements

for f in */SKILL.md; do printf "%-30s %8d\n" "$f" "$(wc -m < "$f")"; done | sort -k2 -rn | head
grep -n '^# /' */SKILL.md | head     # offset of the first skill-specific heading = preamble length

Line-level duplication across all 58 SKILL.md files:

import os
from collections import Counter

sk = {d: open(f"{d}/SKILL.md", encoding="utf-8").read()
      for d in os.listdir(".") if os.path.exists(f"{d}/SKILL.md")}
cnt = Counter()
for t in sk.values():
    for line in {x.strip() for x in t.split("\n") if len(x.strip()) > 25}:
        cnt[line] += 1

common = {l for l, c in cnt.items() if c >= len(sk) * 0.7}
tot = sum(len(t) for t in sk.values())
shared = sum(len(l) for t in sk.values() for l in t.split("\n") if l.strip() in common)
print(f"{len(common)} lines in >={int(len(sk)*0.7)} files; {shared}/{tot} chars = {shared/tot:.1%}")
# -> 404 lines in >=40 files; 1600571/3210730 chars = 49.9%

Suggested fixes

1. Hoist the preamble into a shared include that the generated SKILL.md references instead of
   inlining. Five of the seven offenders drop under 100k immediately (office-hours,
   plan-design-review, plan-devex-review, spec, plan-eng-review); ship (~147k) and
   plan-ceo-review (~110k) would still need real splitting.
2. Emit sections as references/ shards for the heaviest skills, keeping SKILL.md as the
   entry point.
3. Add a size guard to gen:skill-docs that warns (or fails) when a generated SKILL.md
   crosses ~90k chars, so this surfaces at generation time rather than in a downstream tool.

spec is a special case worth a look: 65% of its lines are shared with other skills, so it likely
carries more extractable duplication than the common preamble alone.