docs: substantive tooltips on every INPUT_TYPES input

[laurigates]

Summary

Adds substantive tooltips to every input declaration in INPUT_TYPES across the pack — 573 tooltips touched across 29 files. Coverage went from ~27% substantive descriptions to 100%; every input now has a hover description > 40 chars.

Coverage delta

Status	Before	After
Substantive (> 40 chars)	27%	100%
Terse (≤ 40 chars, often just restating the field name)	36%	0%
MISSING tooltip key	26%	0%
Wired socket with no tooltip	28%	0%

What changed

• 286 in-place tooltip-text expansions — replaced terse hints like "Image to encode" / "end frame" / "mask" with semantic descriptions including shape/range, pre-processing expectations, and special-case behavior.
• 104 tooltip keys inserted into existing options dicts that previously carried default / min / max but no tooltip.
• 183 bare (TYPE,) tuples extended to (TYPE, {"tooltip": "..."}) — mostly wired-socket inputs (CACHEARGS, MULTITALK_EMBEDS, SLGARGS, EXPERIMENTALARGS, FETAARGS, LOOPARGS, UNI3C_EMBEDS, FANTASYTALKING_EMBEDS, …) that now name their producer node on socket hover, so users can discover graph topology by hovering an unconnected input.

Style

• Substantive descriptions (> 40 chars) on every input.
• Wired-bundle sockets name the producer node ("… — connect from WanVideoCacheArgs", etc.).
• Numeric inputs include units / sentinel-value semantics inline (e.g. riflex_freq_index: "disabled when 0, default 6").
• BOOLEAN tooltips lead with the true behavior.
• COMBO tooltips describe the choice axis, not enumerate members (the UI already shows them).

Latent bugs caught + fixed along the way

1. WanVideoAnimateEmbeds.pose_images and .face_images both had tooltip "end frame" — a copy-paste from WanVideoImageToVideoEncode. Replaced with descriptions of the DWPose-driving video and the 512×512 face-identity crop respectively, traced to consumer code around lines 1257 and 1322.
2. NormalizeAudioLoudness.lufs had options key "tool" instead of "tooltip" (typo) — suppressed the existing description from rendering. Renamed.

Implementation notes

• Authored against a local install, then ported to upstream HEAD via an AST-aware patch script that injects/replaces only tooltip strings — no other code is touched, and fields that exist locally but not upstream (or were renamed upstream) are silently skipped.
• Local hot-patches (e.g. the predict_with_cfg zero-init early-exit tuple fix from fix: return 3-tuple from predict_with_cfg use_zero_init early-exit #2011 and the WanVideoScheduler VALIDATE_INPUTS step-range fix from fix(scheduler): reject step ranges that produce empty timestep slices #2012) deliberately did NOT ride along — they're tracked separately.

Verification

• All 29 files compile clean (python -m py_compile).
• Custom-node pack imports in 0.1 s after restart on the test install.
• Spot-checked via /object_info/<NodeClass> HTTP API — tooltips round-trip to the frontend correctly.

Test plan

• Restart ComfyUI and confirm WanVideoWrapper imports clean (no IMPORT FAILED).
• Hover any input in any WanVideo* node to confirm the new tooltip renders.
• Hover a socket like cache_args on WanVideoSampler to confirm socket-hover tooltips work and name the producer.
• (Optional) Spot-check via API: curl -s http://127.0.0.1:8188/object_info/WanVideoSampler | jq '.WanVideoSampler.input.required.cfg[1].tooltip'.

🤖 Generated with Claude Code