[documentation] Trim bloat / duplication

[wlewNV]

FlashDreams docs: trim bloat / duplication

From the same 2026-06-09 docs audit. These docs were AI-generated and carry leftover artifacts, repeated content, and over-explanation. This issue is style/verbosity only — accuracy bugs are in the companion "accuracy fixes" issue.

All paths are repo-relative. Tasks grouped so they can be split/self-assigned.

---

Remove leftover AI artifacts & internal planning (cheap, high-signal)

• docs/source/index.rst:95-96 — remove .. TODO: Vectorize this figure... and .. Figure creation trace: https://chatgpt.com/share/....
• docs/source/developer_guides/inference_pipeline_overview.rst:23 — remove the .. Figure creation trace: https://chatgpt.com/share/... comment.
• integrations/hy_worldplay/README.md:228-268 — move the ~40-line internal "Staging plan" (phases 2b.1–2b.6) to a tracking issue / dev doc; cut to 2-3 lines.
• …/ludus-renderer/ludus_renderer/_cpp/cudaraster/PORT_NOTES.md:186-232 — "__activemask() Audit (Reviewer Add GitHub Actions CI configuration for NVIDIA self-hosted runners #4)" pasted review checklist → tracking issue.
• PORT_NOTES.md:300-344 — "Incomplete Work" / "Pending Phase 2 Audit" inventory → tracking issue (keep the durable parts: tiebreaker rationale :15-113, "deviations from upstream" :117-185).
• …/ludus-renderer/docs/interop-optimization-roadmap.md — the whole 155-line optimization wishlist → issue tracker; keep at most a short "perf characteristics + how to run benchmark_renderer.py" note.

Done when: rg -n "chatgpt.com" docs/ integrations/ returns nothing and no published doc contains phase-tracking checklists or reviewer to-do lists.

Stop repeating long prompts in model pages (good first issue)

Replace 2nd/3rd copies of the long demo prompt with a short one (e.g. "A cat surfing."); drop --prompt where it's already the runner default.

• docs/source/models/self_forcing.rst:68,95 (same prompt 2×)
• docs/source/models/wan21.rst:55,66,76 (3×)
• docs/source/models/causal_forcing.rst:67,79,90 (3×)
• docs/source/models/causal_wan22.rst:54,61,73 (drop one near-duplicate single-GPU block)
• docs/source/models/cosmos_predict2.rst:74,84,93,129,138 (5×; it's the default prompt runner.py:46, so drop --prompt from run blocks entirely)

Trim over-length agent skills

(skills/README.md:40 sets a ~500-line guideline + "push long-form into sibling files". Coordinate with the skills accuracy fixes to avoid churn.)

• skills/maintaining-oss-state/SKILL.md (630 lines): cut TL;DR :22-63 (dup of §1–8), §12 checklist :551-616, §13 table :618-630 (dup of :67-77); move long-form into a sibling reference.md.
• skills/flashdreams-integrations/SKILL.md (423 lines): merge §8 pitfalls :372-424 (restate §2–5); dedupe "two encoder slots" (:96-103,172-177,363-366,379-381) and the KV-cache/compile guidance (:307-321 vs :405-412).
• skills/python-docstring-style/SKILL.md (406 lines): merge :323-383 with :398-407; consolidate the "don't restate the next line" message (:250-269,271-276,336,396); soften :320 (the repo does register Phases:/Per-step usage: via napoleon_custom_sections, conf.py:129-135).

Trim verbose docs & dedupe shared content

OmniDreams

• docs/source/models/omnidreams.rst:135-294 (~160 lines) re-duplicates the 608-line interactive_drive/README.md → reduce to ~10-line overview + link.
• docs/source/models/omnidreams.rst:221-264 steering-wheel/FFB deep-dive → move to the subpackage README; :296-362 WebRTC section duplicates integrations/omnidreams/README.md:72-106 → condense + link.
• integrations/omnidreams/omnidreams/interactive_drive/README.md (608 lines): collapse OOB respawn :501-557, FFB/vendor-driver :275-362, synthetic-scene internals :559-584, and browser-HUD internals :399-424 to their tables + a sentence each.

FlashVSR / tests / API

• integrations/flashvsr/README.md (337 lines): the 128-alignment explanation is tripled (:65-73,211-220,268-275) → keep one; "Builder knobs" :263-289 duplicates the build_flashvsr_v1_1 docstring; condense the Files map :291-304; collapse 3 near-duplicate gRPC invocations :116-170.
• tests/README.md:33-148 — Omnidreams quality section repeats the example UUID 3× (:50,69,83) and the "extract lower half" note twice (:88-94,121-125); state once, consider relocating the runbook.
• docs/source/api/infra.rst:113-166 — Encoder/Decoder bullet prose duplicates the module + class docstrings autoclass already renders; trim to a one-line lead-in.

Cross-file dedupe & orphans

• Dedupe the near-identical "Install (workspace member)" paragraph + HF_TOKEN/HF_HOME block across integrations/{self_forcing,wan21,causal_forcing,fastvideo_causal_wan22,lingbot}/README.md (lingbot states HF/disk reqs twice: :50-61 and :169-180). integrations/wan22/README.md is the concise model; consider a shared include.
• docs/source/developer_guides/usage_patterns.rst and interactive_serving.rst are :orphan: + commented out of the toctrees (index.rst:289-291) and duplicate index.rst/installation.rst — publish or delete.
• README.md:18-23 — trim the marketing laundry-list ("…across gaming, autonomous vehicles, robotics, …, and more").

[mvanhorn]

Implemented in #341 - docs: Remove leftover AI artifacts and dedupe repeated demo prompts in model doc.

[jarcherNV]

I checked this again. #341 was closed unmerged, and the cleanup it attempted does not appear to be present on current main.

I’ll take this over from here. The first follow-up should probably be a smaller docs PR that revives the useful parts of #341 against current main: removing the remaining AI artifact comments and deduping the repeated long demo prompts. After that we can either continue through the remaining checklist items here or split the larger docs-trimming work into smaller follow-up issues.