feat(mofa-slides): migrate 670→51 line SKILL.md to discovery+docs/ tree (PR-D2)

[ymote]

Context

mofa-slides is the largest skill on the fleet — its 670-line SKILL.md is the bulk of every fleet turn's standing token cost. This PR ports it to the discovery+docs/ tree model so the LLM reads reference content on-demand via manifest hints instead of paying for it up front on every turn.

Prerequisites (octos): PR-A #1327, PR-B #1333, PR-C #1339, PR-D1 #1340.

This is the single biggest token-cost recovery in the fleet — ~11K tokens/turn savings post-PR-E once the host stops auto-injecting SKILL.md by default (skill_md_auto_inject: false).

Before / after

	Before	After
SKILL.md	670 lines	51 lines
Reference docs	—	5 files under docs/
Manifest	no discovery	discovery.summary + 8 hints
skill_md_auto_inject	(implicit true)	false
Version	0.6.1 (head was 0.6.0)	0.7.0

Load-bearing anti-patterns preserved in the trimmed SKILL.md

The full 51-line SKILL.md keeps only what cannot be inferred from tool descriptions:

• Output-path rebind rules (skill-output/ auto-prefix; no /tmp/).
• Workspace-contract truth (check_background_tasks + check_workspace_contract are authoritative — not chat history).
• Interactive flow one-liner.
• The four load-bearing anti-patterns:
  1. Do NOT memorize the style catalog — call mofa_list_styles. The binary errors out (no silent fallback) on unknown names.
  2. Language match end-to-end. Chinese deck → Chinese VQA + Chinese narratives + Chinese quoted text. No mixing. Highest-impact rule for Chinese decks. The runtime no longer auto-appends an English clamp.
  3. Mode 2 (clean bg + manual texts) is the default for "editable PPT" — NOT Mode 3 (auto_layout VQA). Mode 3 is slow, expensive, needs human cleanup.
  4. Every deck is its own self-contained script with a per-deck const VQA block at the top, spliced into every slides[i].prompt via ${VQA}. Write script.js to disk first; invoke with input: "<path>".
• Pointer to docs/ + the manifest's discovery.hints.

Migrated content → mofa-slides/docs/*.md

5 new self-contained docs, each readable in isolation:

• docs/modes.md — Mode 1/2/3/4 decision tree, when NOT to use auto_layout, costs of enabling it, anti-leak rules, full 4-phase pipeline description.
• docs/cc-ppt-authoring.md — Required deck-script structure (top constants, const VQA, module.exports), mandatory rules, full Chinese + English minimal worked examples.
• docs/custom-styles.md — Inline prompt overrides AND full-TOML custom style authoring (PR feat(slides): document full-TOML custom style authoring #72 content: workspace save path, TOML schema, workflow, the three "style not found" failure modes).
• docs/worked-examples.md — Full 5-slide editable business deck (Mode 2 with texts, runs, fill, margin, lineSpacing), Mode 3 / Mode 4 / reference-image JSON shapes, slide-canvas layout reference (16:9, inch positions), complete TextOverlay + TextRun schemas.
• docs/style-prompts.md — Inline style-prompt quick-reference table (14 vibes: Art Deco, Bauhaus, Glassmorphism, 国潮, 水墨, 敦煌, 青花瓷, …), built-in style category cheatsheet (verify with mofa_list_styles), API modes, timing/timeout budget, CLI flags, config, OOXML + pptx-scripts/ utilities for editing existing decks.

Discovery hints (8 / 8 max, one per concern)

Each hint names a trigger condition + the doc to read:

1. Before picking a style → call mofa_list_styles (never recite).
2. Choosing Mode 1/2/3/4 → docs/modes.md.
3. Authoring a deck (cc-ppt structure) → docs/cc-ppt-authoring.md.
4. Inline overrides + full-TOML custom styles → docs/custom-styles.md.
5. Long worked examples + layout reference + schemas → docs/worked-examples.md.
6. Style-prompt snippets, API modes, CLI, PPTX-editing → docs/style-prompts.md.
7. Always invoke via input: "<path>" with ${VQA} spliced (not inline slides:).
8. Language match (Chinese stays Chinese; English clamp removed).

Untouched

• Tool descriptions in manifest.json (mofa_list_styles, mofa_slides) — identical.
• mofa-slides/styles/ (20 TOML files) — untouched.
• mofa-slides/pptx-scripts/ — untouched.
• mofa-slides/ooxml/ — untouched.
• mofa-slides/USER_GUIDE.md and the pre-existing docs/00-SKILL-完整制作指南.md — unchanged.

Test plan

• manifest.json parses as valid JSON.
• SKILL.md is 51 lines (≤ 80 cap).
• All 5 paths referenced by discovery.hints exist under mofa-slides/docs/.
• All 4 load-bearing anti-patterns are present verbatim in the trimmed SKILL.md.
• styles/, pptx-scripts/, ooxml/ subdirs untouched.
• Fleet soak (post-merge + PR-E in octos): verify slides authoring still works end-to-end on Chinese + English decks with the LLM reading docs/ on demand.