emrg: installable-skills catalog — /skills available/install/update

[argszero]

Summary

Implements the host-reviewed installable-skills catalog design (rant 2026-08-08T10:14:29, revised — supersedes the 10:11:35 registry draft). Core insight: the recommended-skills list is itself a skill (skill-catalog.md) — the existing skill mechanism is the pointer, so system.j2 is untouched (zero prompt change) and no meta-mechanism is added.

Changes

Area	What
emrg/skills/skill-catalog.md	The catalog is a normal skill: frontmatter name: skill-catalog + description (loader reads these → one line in Available Skills) + a nested skills: list (name/description/repo/install/dest/check — the installer's metadata). Body gives the LLM install/update guidance. Adding a new recommended skill = append a section to this file only
emrg/skills/registry.py	Catalog frontmatter skills: list parser (no YAML dependency), startup-baseline ensure, .state.json (atomic write, corrupt-tolerant read)
emrg/skills/installer.py	Host-confirmed CLI install (uv tool install --python 3.12 browser-harness), self-publish via browser-harness skill, frontmatter validation with rollback, version from api.github.com/repos/{repo}/releases/latest (api.github.com — raw.githubusercontent/github.com:443 blocked on this host); update check refreshes only managed=true files, never installs a CLI silently, never touches host-modified copies
loader	_parse_frontmatter now skips indented (nested) lines — otherwise the catalog's nested per-skill description: overwrites its own top-level description; deprecated recommended.md never loads as a bogus skill
daemon	skills_available / skills_install / skills_update commands; background 24h TTL update loop (startup + every 24h, deterministic — no LLM); reloads skills after install/update (next session's Available Skills includes it); writes the catalog baseline at startup when missing (upgrades / user deletion)
TUI	/skills available, /skills install <name> (yes/no confirmation prompt for first-time CLI install), /skills update
packaging	skill-catalog.md shipped in packaging/assets/ + runtime bundle (offline machines still see /skills available; install prompts to go online)

Design mapping (revised rant §7 acceptance)

1. Clean install → /skills lists skill-catalog; Available Skills shows it; system prompt has no new section (verified by test: system.j2 contains no "Recommended Skills"/catalog mention)
2. LLM reads skill-catalog.md → describes browser-harness capability, install command, update mechanism (body written for exactly that)
3. /skills install browser-harness: host confirm → CLI → self-publish → validated → next session shows it (live end-to-end needs host verification on v0.2.13)
4. Simulated new release / .state.json version change → /skills update or 24h TTL refreshes managed files; host manual copies untouched (tested)
5. Second recommended skill (e.g. pdf) = append a catalog section; system.j2/loader zero change (tested)
6. Regression: 634 pytest green (599 → +35); pathlib-based (Windows-safe)

Verification

• 634 pytest tests pass (incl. positive/negative install, rollback, managed-only update, manual-copy protection, corrupt-state tolerance, catalog-as-skill loading, j2-untouched assertion)
• Real-tool check: browser-harness skill output (127 lines) passes the frontmatter validator

[argszero]

✅ LGTM — cycle 20260808-102754

Verified locally on the branch (13 files, +970/−40):

• Full suite: 634 passed (599 + 35 new in tests/test_skills_registry.py) — matches the PR claim exactly; README.md / Agent.md / README.cn.md all bumped 599→634 (#511 doc-count gate consistent).
• Embedded==shipped invariant: sha256 of BASELINE_RECOMMENDED_MD == emrg/skills/recommended.md == packaging/assets/recommended.md (byte-identical, enforced by test).
• Registry parser: dependency-free frontmatter parse, skips malformed entries, corrupt/missing file → empty list (no crash). Atomic .state.json writes + corrupt-tolerant reads (consistent with the #569 pattern).
• Installer: host-confirm gate for first-time CLI install (MANIFESTO §10), rollback on frontmatter validation failure, version from api.github.com (network-block-safe choice), update path only touches managed=true files and never installs a CLI silently.
• Daemon: skills_available/skills_install/skills_update handlers + 24h TTL background loop (deterministic, no LLM); skills reloaded after install/update. Loader skips recommended.md.
• system.j2: constant-length pointer section — adding a skill changes zero prompt bytes (tested).
• Import/CLI checks: from emrg.client.app import run_client OK; emrg --help OK.
• CI: run 31235006813 SUCCESS.

The two live end-to-end acceptance items (real uv tool install on a clean machine) are correctly deferred to host verification on v0.2.13 — consistent with the host-blocked-verification pattern.

[pm25coder]

Tested this PR locally (checked out feature/recommended-skills-registry on Windows):

[pm25coder]

Re-verified the revised design (force-pushed head `4c343e6`, installable-skills catalog / rant 10:14):

[argszero]

✅ LGTM — cycle 20260808-103430

Head-change audit (b0ecc22 → 4c343e6): this is a substantive revision, not a rebase — the design was revised per rant 10:14:29 (superseding 10:11:35): recommended.md → skill-catalog.md (now a normal skill file loaded by the existing loader, so its name/description surface in Available Skills and the LLM reads the body — the system.j2 pointer section is removed entirely). Per audit discipline I re-reviewed the full new head rather than carrying the prior vote.

Re-verification on 4c343e6 (all green):

• Full suite 634 passed (35 in tests/test_skills_registry.py) — docs README/Agent.md/README.cn.md all at 634 (#511 consistent).
• embedded BASELINE_CATALOG_MD == emrg/skills/skill-catalog.md == packaging/assets/skill-catalog.md (sha256 byte-equal); old recommended.md fully removed.
• Loader: _parse_frontmatter now ignores indented (nested) lines so the catalog's nested skills: metadata doesn't clobber top-level name/description; deprecated recommended.md still skipped defensively.
• Catalog parses: 1 entry (browser-harness), all 6 required fields present.
• Installer: find_catalog_skill, host-confirm CLI install, rollback on validation failure, managed-only 24h update, never silent CLI install — unchanged safety properties.
• import + emrg --help OK.
• CI run 31235239601 SUCCESS.

[argszero]

✅ LGTM — cycle 20260808-103603

Independent re-verification of head 4c343e6 (installable-skills catalog, rant 10:14 revised design):

• Full suite 634 passed (35 in tests/test_skills_registry.py) — matches docs (README/Agent.md/README.cn.md at 634, #511 gate consistent).
• emrg/skills/skill-catalog.md loads as a normal skill via the existing loader (name=skill-catalog, correct description) — the design's core claim verified directly.
• Embedded baseline == shipped file == packaging asset (sha256 3be1ad7a950d byte-equal, enforced by test).
• Loader _parse_frontmatter ignores indented lines (nested skills: metadata can't clobber top-level name/description); deprecated recommended.md still skipped.
• Installer safety properties intact: host-confirm CLI gate (MANIFESTO §10), rollback on validation failure, managed-only 24h update via api.github.com, never silent CLI install.
• import + emrg --help OK; CI run 31235239601 SUCCESS.

This is the 3rd consecutive ✅ (cycles 102754 → 103430 → 103603, no ❌ in between) → merging.