refactor(skills): retire the context-manager scripts and align every Skill name with its directory

[jiunbae]

Two commits. The first was the original scope; the second closes a naming split
the first one surfaced.

---

1 — context-manager loses its scripts and leaves core (e75ecc2)

find_context.py presented itself as a weighted relevance ranker. Measured
against a real context/ directory:

$ find_context.py --context-dir IaC/context --keywords deploy onboarding \
    --task-type implementation
1. context/README.md                              0.700   ← pinned, always first
2. context/operations/service-onboarding.md       0.209   ← the correct answer
3. context/architecture/sealed-secrets.md         0.204
4. context/architecture/security-audit-report.md  0.204
5. context/architecture/nfs-storage.md            0.200   ← unrelated

The correct document beat an unrelated one by 0.009. The causes are structural:

• Bodies are never read. score_keyword_match looks at the filename stem
  and the parent directory name only, so any term that lives in the text —
  the normal case — scores zero.
• README.md is hardcoded to 0.7, so it always ranks first.
• The 10% recency weight is inverted in effect: a document untouched for
  90 days scores zero, penalising the stable, canonical procedures.
• score_path_match compares code path segments against context category
  names; the vocabularies rarely intersect.
• min_score=0.1 is met by recency alone.

grep -ril over document bodies is strictly more accurate, so SKILL.md now
says that. update_context.py goes too — the skill does not write documents on
its own initiative, and its mkdir contradicted the instruction not to create
the directory. references/context_patterns.md (524 lines) duplicated
static/CONTEXT.md and cited layouts that no longer exist.

The core profile is defined in five places that had to stay in agreement:
profiles.yml, install.sh (array + help text), install.ps1, README.md,
README_ko.md. The skill stays in full and installable by path.

2 — every Skill declares the name a reader types (30d992a)

e529d78 renamed 22 Skills' name frontmatter to gerund phrases. A host lists
a Skill by its directory, so managing-vault-secrets on
integrations/vault-secrets was a second name nobody could type.

Not cosmetic: air-workbench's catalog keys items on the frontmatter name, so the
catalog and the host disagreed about what a Skill is called — catalog.test.mjs
asserted analyzing-business-model for a directory named bm-analyzer.

Renaming the directories was the alternative and is worse: it breaks every
published --from jiunbae/agent-skills/<group>/<skill> path and the CHANGELOG's
own history. So the declaration moved, in 22 one-line edits.

tests/core.test.mjs gains a guard that fails with the offending path on any
future divergence — verified by reintroducing a mismatch and watching it fail.
The rule is documented in both READMEs.

Test fallout, explained

• Frozen security-auditor node ids changed. A node id hashes the heading's
  byte offset; the frontmatter line is one byte shorter. Titles and graph shape
  are unchanged — recomputed and updated.
• The playwright search fixtures are synthetic and keep testing a divergent
  name, which an author can still write. Only their comments changed: they no
  longer claim the repository contains such a Skill.

Verification

• bash -n install.sh; --help lists the 7 remaining core skills.
• air-workbench: 355 tests pass (354 before, +1 guard).
• Drift guard negative-tested.
• No dangling references to removed files or retired names.

🤖 Generated with Claude Code