Media Generator

Marketplace listing update: added the launch explainer film (`media.demo_video_url`). No code, schema, copy or guidance changes; hero and screenshots unchanged from 1.5.3.

Packaging + visibility fix. No change to the skill's behaviour.

- 🔴 **The package was INVISIBLE.** `skills[0].visibility` was `internal` while `tier` was `beta`,
  so the registry reported it published at 100/100 HEALTHY while **no brand could install it** —
  for 34 days, since 1.5.2. Now `public`.
- **`scripts/verify-coordination.mjs` crashed on import in every prior release.** It imported
  `'../../../scripts/_verify-coordination-shared.mjs'`, which from this bundle's root-level
  `scripts/` dir resolves *outside the bundle entirely*, and the helper was never staged beside it.
  The gate died with `ERR_MODULE_NOT_FOUND` on every run and verified nothing, silently. The
  wrapper is now self-resolving and the helper is vendored.
- **Removed internal workbench material that was shipping to customers** — `BRIEF.md`, `tests/`,
  `source/`, `propagation/`, `marketing/` (27 files). Customer-facing content is unchanged.

**SPEC-3 asset kit for the marketplace card. No skill-behavior change.**
- Added top-level `media:` block — hero (studio-orchestration metaphor, 16:9) + 3 screenshots (generate→persist→reference loop · brain/hands peer split · AI-slop-vs-brand-grade), all dark-board chalk (#131313 / white / gold #EEBF01), produced via SpiderGate `kie/nano-banana-2`, durable `media.spideriq.ai` URLs.
- Refreshed `marketing.long_description` to first-person Martin (Problem → Solution → worked example → peer chain); added 5 verb-led `marketing.features[]` with emoji icons; author → Martin Shein.
- Content/behavior identical to 1.5.1.

**Bundle-shape migration to nested co-location. No content change from 1.5.0.**
- Moved SKILL.md + schema.yaml into `skills/media-generator/` (co-located), repointed `package.yaml` `skills[].schema` → `./skills/media-generator/schema.yaml`. This is the shape `verify-packaging.mjs` mandates so `skill_read` resolves SKILL.md from the schema's own dir.
- Root `SKILL.md` + `client/schema.yaml` retained; content identical to the nested copies.
- **GATED** on `@di-atomic/outbound-engine@2.2.0` (card `8f0144d6-…`) proving `skill_read` resolves live from the nested shape. PASSES → ship 1.5.1 nested. FAILS → do NOT ship; 1.5.0 flat stays live.

**Wave-3 enrich from deployed-agent experience (Lola @ VayaPin brand 4).** Additive — no scope change, no restructure.
- **SKILL.md:** LOUD guidance-only banner at top — names the exact failure (`method 'generateImage' has no parsable 'http' clause`) and the exact peer to call instead (`generate-media.list` → `generate-media.generate` / `gate_media_models` → `gate_media_generate`). Stops agents from `skill_call`-ing the documented-but-uncallable methods. (Harvest I1 + G1.)
- **learnings/2026-07-19-methods-index-lies-guidance-only** — the method signatures are a documentation/routing surface, not executable endpoints; the callable engine is the peer @spideriq/gateway-skills. (G1, confirmed ×3 Lola sessions on v1.3.2 + v1.4.0.)
- **learnings/2026-07-19-read-model-inputs-first** — undeclared params are silently dropped; discover model inputs (list) + validate BEFORE generate. Casing matters (`aspectRatio` vs `aspect_ratio` — wrong one is dropped, not coerced). (G3.)
- **recipes/media-asset-lifecycle** — generate → persist (CONDITIONAL) → reference. Distinguishes already-SpiderMedia-stored gate output (skip persist) from a raw provider/kie.* URL that bypassed the gate (run upload-host-media). (R1.)
- **recipes/image-prompt-guardrails** — no fake logos / no readable fake text / no named characters — the safety pass before every image reaches the engine. (R3.)
- Coordination cross-refs (v2.2) already present; no v2.4 upstream (stays guidance-only — the engine is a peer). Versions bumped 1.4.0 → 1.5.0.

**Adopt the schema-aware media backend (@spideriq/generate-media / SpiderGate /media/generations).** Execution routes through gate_media_generate (discover inputs via gate_media_models first; pre-flight validate-media-params). Added peer @spideriq/gateway-skills >=0.5.0. Persistence simplified (gate returns SpiderMedia-stored URLs; upload-host-media redundant for them). references/spidergate-routing.md rewritten; Mode A/B demoted to legacy. Runtime prereq: SpiderGate MCP (@spideriq/mcp or @spideriq/mcp-gate).

**Fix: `skill_read` 404 (guidance was unreachable).** v1.3.0's `skills: []` made the package a pure meta-package (install-expands-to-peers) that registers NO skill id — so `skill_read(media-generator)` had nothing to resolve and 404'd (the bug OPVS flagged across di-atomic guidance skills converted under the 2026-06-17 directive).
- Restored `skills: [{ id: media-generator, schema: ./client/schema.yaml, visibility: internal, entitlements: [] }]` + the `client/schema.yaml` — the readable guidance-only shape (matches the working `@di-atomic/outbound-engine` + packaging.md §1: guidance-only = a skills entry + schema, no backend).
- Kept the 1.3.0 peer_dependencies + README + CHANGELOG + docs.skill_guide. Still no real backend — agent skill_reads the id, composes the real SpiderGate calls itself.

**Converted to guidance-only meta-package shape** (per OPVS packaging spec §4 + the @di-atomic/copy-engine@0.4.1 lighthouse).
- `skills: []` + `peer_dependencies` (4 OPVS coordination skills + copy-engine + @spideriq/media-skills). Agents reach the skill via `skill_read`.
- Removed the method-bearing `client/schema.yaml` + `tools-manifest.json` from the bundle (media-generator has no backend of its own; the agent composes the real SpiderGate calls itself). Preserved at `source/legacy/client-v1.2.1/` for provenance.
- `provider_config` kept as the OPVS_API_URL/OPVS_PAT placeholder (validator requirement; no method calls it).
- Added `README.md` (marketplace listing) + this `CHANGELOG.md`; `docs.skill_guide` → `./SKILL.md`.
- Rationale: SpiderGate smoke test (2026-06-17) confirmed the gate is real (`/v1/models` 200, `/v1/images/generations` 401 not 404), but that is a separate service — media-generator exposes no methods to proxy, so guidance-only is the runtime-correct shape.

- Canonical shared-helper `scripts/verify-coordination.mjs` (matches retrofit-card template). verify-coordination PASS.

**HeyGen Hyperframes lessons.**
- `verify-brand-consistency` now MEASURES rendered pixels (sharp via `package-loader.mjs`); UNVERIFIED (never auto-PASS) when it can't measure — closes a fudge hole.
- New `verify-all.mjs` client-grade gate (one ship/don't-ship verdict).
- Deterministic helpers: `prompt-prep.mjs` (800-cap + per-model aspect quirk), `compose-timeline.mjs` (keyframe math).
- Monoculture tags on 5 creative registry blueprints ("calibration, not a formula").
- HARD-GATE names anti-defaults; added "build exactly what was asked" rule; tiered MUST/should/prefer.

- **SpiderMedia persistence** — every generated asset persisted via `upload-host-media` → durable `media.spideriq.ai` CDN URLs (provider URLs expire).
- **v2.2 coordination** cross-references declared (4 OPVS surfaces); `verify-coordination` gate.

- Initial release. 6-method workflow, 8 recipes, 30 registry items, 11 learnings, 3 enforcement scripts, SpiderGate Mode A/B routing. Built from 24 production workflows. 100/100 security score.