Problem: Build a producer-ready `Totally Spies` proposal and an executable
owned-asset validation repo grounded in official franchise marketing references,
while keeping all benchmark critique fully internal.

Approach:
- Use verified official assets from the official `Totally Spies!` channel and
  Banijay rollout context as the marketing reference set.
- Treat `Brand New Gadgets!` and `Best Gadgets!` as the short-form target stack
  and the `Season 7 Trailer` as the campaign master benchmark.
- Use the Banijay-linked Vimeo trailer as the current downloadable official
  source for frame-level analysis.
- Translate the benchmark findings into the repo's validation shot plan, asset
  request pack, and working shot priorities.
- Keep demo critique and generic benchmark comparisons out of external wording.

Todos:
- [x] Identify the official reference sources and rollout context.
- [x] Verify representative official assets from primary sources.
- [x] Derive the working shot taxonomy from those assets.
- [x] Create a revised draft with taxonomy-driven framing.
- [x] Materially restructure the draft around a monthly content menu and modular campaign families.
- [x] Shift the draft to an action-oriented execution voice.
- [x] Create a standalone local repo for the project and add the relevant working materials.
- [x] Expand the repo with the full media set, LFS rules, and a native devenv scaffold.
- [x] Explore the project baseline and define the owned-asset validation structure.
- [x] Scaffold owned-episode ingest and deterministic reference extraction.
- [x] Create the internal asset request pack and validation trackers.
- [x] Rank the official benchmark stack.
- [x] Download an official trailer mirror and extract benchmark frames.
- [x] Write the official benchmark frame-analysis note from the Vimeo capture.
- [x] Optional next step: tighten the proposal further for either commercial emphasis or technical emphasis.
- [x] Optional next step: obtain local or alternate-official copies of the short-form gadget clips if frame-level capture is still needed beyond the trailer.
- [x] Download official S7 YouTube corpus (104 videos — pure-s7 + mixed-1-7).
- [x] Build fandom wiki database and correct canonical character/gadget/location data.
- [x] Revise proposal scope to image-first cycle ($12,500 / EUR 7.5K budget).
- [x] Conduct S7 animation style deep research (15 research docs, 3 benchmark packs).
- [x] Run multi-model pack evaluations (qwen2.5vl:7b, gemma3:12b, gemma4:26b).
- [x] Confirm Qwen3-VL 235B cloud as primary visual analysis model (Gemma4 26B local non-functional).

Notes:
- Verified primary references include the official `Totally Spies!` YouTube
  channel, the official `Totally Spies! Season 7 Trailer`, the `Top Secret
  Mission` music video, `Brand New Gadgets!`, `Best Gadgets!`, and Banijay's
  ITVX / POP rollout note.
- The official Banijay Vimeo mirror used for current frame analysis is:
  `https://vimeo.com/953985089`
- Soft or secondary sources can help with context, but reposts and archive
  channels such as `CNArchives` should not be treated as primary marketing
  references.
- Current external draft: `r-2510603559359686682`.
- Current Laurent reply draft: `r-7082543567781997488` (image-first scope, $12,500).
- Standalone repo created: `/home/mnm/workspaces/totally-spies-cultshot`
- Current internal shot taxonomy:
  - trio-led setup shot
  - gadget or prop reveal
  - location or HQ beauty shot
  - villain or problem beat
  - fashion or suit beat
  - hero or dialogue closeup
  - short platform-specific cutdown
- Material rewrite applied:
  - lead frame changed from generic pipeline to repeatable monthly marketing system
  - added concrete monthly output menu: anchor promo cut, platform cutdowns, reusable modular beats
  - added campaign operating model showing how briefs become shot families, batches, hero finishing, and editorial versions
  - commercial frame now explicitly maps price to a mixed package of deliverables rather than an abstract volume bucket
- Tone rewrite applied:
  - removed hedge language such as `I think`, `I'd want`, and `I'd suggest`
  - switched to direct execution phrasing: what Cultshot will build, what the team will provide, how validation will run, and what the next step is
  - closing now points to the validation package as the practical next move without sounding pushy
- Repo and benchmark expansion applied:
  - repo includes the owned-asset manifest, validation shot plan, metrics tracker, asset request pack, ingest tooling, and extraction tooling
  - official benchmark ranking lives in `docs/research/official-benchmark-shortlist.md`
  - official frame-level findings now live in `docs/research/official-benchmark-frame-analysis.md`
  - `nix run nixpkgs#yt-dlp` works, but official YouTube retrieval remains blocked by login / anti-bot enforcement
  - the Banijay Vimeo trailer mirror was downloaded and reviewed through timed and scene-change contact sheets
  - the official trailer now also lives in the repo at `materials/source-media/official-totally-spies-season-7-trailer.mp4`
  - the repo-local analysis bundle now lives in `materials/benchmark/official-vimeo-trailer/`
- First-cycle execution brief created:
  - `docs/strategy/top-3-validation-shot-brief.md`
  - converts `TSV-001`, `TSV-002`, and `TSV-005` into a concrete operating spec
    with framing, action, owned-input requirements, hard failure conditions, and
    fill-before-run fields
- First-cycle handoff package created:
  - `docs/internal/top-3-validation-delivery-request.md`
  - rewrites the first ask to Laurent and Paul around the top 3 validations
  - seeds `docs/internal/reference-extraction-requests.csv` with `draft` rows
    aligned to `TSV-001`, `TSV-002`, and `TSV-005`
- Current validation priority order after the official trailer pass:
  1. `TSV-001` trio-led setup
  2. `TSV-002` gadget reveal
  3. `TSV-005` hero/dialogue closeup
  4. `TSV-003` HQ beauty
  5. `TSV-004` villain/threat beat
- Repo scaffold applied:
  - initialized a local git repo on branch `main`
  - added `.editorconfig`, `.gitignore`, and `README.md`
  - added proposal, strategy, research, and internal benchmark docs under `docs/`
  - copied the WhatsApp context summaries and benchmark stills into `materials/`
- Session b032b830 applied:
  - downloaded official S7 YouTube corpus (104 videos, ~11.6GB, via LFS)
  - built fandom wiki database (`data/fandom_wiki/`) with canonical character data
  - corrected: Shmagi (not Schamagy), Jerry as WOOHP founder, Britney as protagonist
  - revised proposal to image-first $12,500 scope (Banijay EUR 7.5K/month confirmed)
  - created scope comparison doc and Laurent reply (draft r-7082543567781997488)
  - conducted S7 animation deep research: digital 2D cutout, restrained motion regime
  - key finding: S7 style not reliably distinct from peer rigged 2D shows — identity fine-tuning is non-optional
  - model finding: Qwen3-VL 235B cloud is the only functional VLM for this task (Gemma4 26B local fails)
  - built youtube-s7-validation benchmark set: 3 benchmark packs, multi-model evals
- Pending: Gmail draft r-7975363520388870160 still has old Shmagi spelling; 478-frame stratified Qwen3-VL run not yet complete; transcript overrides from session dc10e698 still pending; 7 stale docs still to delete