feat: target Blender 5.2 LTS and witness NodesModifier input writes

[TMHSDigital]

Summary

• Retarget the repo to Blender 5.2 LTS primary, 5.1 prior stable, 4.5 LTS fallback. Remove the stale “5.2 sweep planned July 2026” strings.
• Close the gallery witness gap: new examples/gn-modifier-inputs/ writes a Geometry Nodes Float input through the version-appropriate path, reads it back, and asserts evaluated Z-extent == 1 / 2 / 3. The 5.1 dict form is still taught as correct on 4.5/5.1; 5.2 uses mod.properties.inputs.Socket_1.value and the dict form raises TypeError rather than silently no-opping.
• PR smoke matrix is 5.2 + 4.5 LTS. 5.1 runs on the weekly cron only (one workflow, conditional matrix).

This first CI run includes an intentional 5.2 canary (exit 1 after Print Blender version). The 5.2 job must go red. A no-op download is indistinguishable from a pass. After the red run I will revert the canary and wait for green.

Witness (the point of the PR)

Shared GN tree, identifier Socket_1, three modifier copies, scales 1 / 2 / 3. Closed form: 1 m cube × S, lift S/2, evaluated Z-extent == S, zmin == 0.

Falsification — live-run-proven

Probe	Binary	Exit	Error
--api auto	.scratch/blender-5.2.1-windows-x64/blender.exe reports Blender 5.2.1 LTS (hash 9e2066aef7ef)	0	extents 1.000 / 2.000 / 3.000
--api dict	same 5.2.1 LTS	5	TypeError: bpy_struct[key] = val: id properties not supported for this type
--api auto	.scratch/blender-4.5.11-windows-x64/blender.exe reports Blender 4.5.11 LTS (hash 4db51e9d1e1e)	0	extents 1.000 / 2.000 / 3.000
--api rna	same 4.5.11 LTS	5	AttributeError: 'NodesModifier' object has no attribute 'properties'
--api auto	.scratch/blender-5.1.2-windows-x64/blender.exe reports Blender 5.1.2 (hash ec6e62d40fa9)	0	dict path, extents 1/2/3

If either form were hardcoded, one of those red runs would have been green. They were not.

CI download URL — live-run-proven

Python urllib against https://download.blender.org/release/Blender5.2/ (User-Agent set; bare urllib got 403). Matches in the listing: blender-5.2.0-linux-x64.tar.xz, blender-5.2.1-linux-x64.tar.xz. sort -V | tail -1 equivalent resolves:

https://download.blender.org/release/Blender5.2/blender-5.2.1-linux-x64.tar.xz

Framing / contact sheet

• Render path on 5.2.1 LTS Cycles 32 spp, check_framing live-run-proven: fill y=0.806 (band 0.70–0.90), margins min=0.028, no edge touch.
• Contact sheet: docs/gallery/contact-sheets/gn-modifier-inputs-contact-sheet.webp (candidate | armature-bend | damped-track-aim | bmesh-gear).
• Mean Rec.709 luma — live-run-proven: candidate 61.6, armature-bend 71.08, damped-track-aim 43.9, bmesh-gear 77.71. Candidate sits in the calibration cluster; stage is dark; 1/2/3 stairs are thumbnail-legible. Not an asset-type example (no asset sheet).
• Hero webp 9924 bytes (Pillow RGB quality 85). Generated <img alt> on the card is the full teaches string ending mod.properties.inputs.Socket_1.value. — no teaches.split(".")[0] truncation, no <identifier>. Witnesses callout matches witnessesFix character-for-character — inspection-only of generated HTML after python scripts/build_gallery.py.

Targeting / matrix — mixed

• Skill geometry-nodes-python now shows both dict and RNA paths, TypeError vs AttributeError — inspection-only of the skill text; behavior proven by the example runs above.
• blender-smoke.yml conditional matrix: schedule → ["5.2","5.1","4.5"], else ["5.2","4.5"] — inspection-only until this PR’s jobs run.
• 5.2 canary is in this SHA on purpose.

Untested-teaching surface (count only, no fixes)

Skills/snippets that teach a load-bearing API no example asserts:

1. NodeSocketBundle / Combine+Separate Bundle
2. For Each Element zone
3. Repeat Zone (GeometryNodeRepeatInput/Output)
4. bpy.app.handlers.exit_pre (5.1)
5. snippet usd-export-evaluation-mode.py
6. snippet cross-version-property-delete.py

Template-only (templates exist, no gallery example): addon-scaffolding / register_classes_factory, ui-panels, custom-properties / PointerProperty, bl-info-migration.

6 untested skill/snippet contracts + 4 template-only. foreach_get has only a weak token hit. This PR closed the NodesModifier input gap; the rest stay for later.

Test plan

• 5.2 smoke job goes red on the canary (this SHA). Record the job URL and the resolved Downloading https://…blender-5.2.1-linux-x64.tar.xz line.
• Revert canary, push.
• 5.2 + 4.5 smoke green on the post-canary SHA. Confirm binary versions in the logs.
• Validate + drift-check green. Socket Security Project Report + Pull Request Alerts green (do not merge while pending).
• Squash-merge with feat: title, delete branch.
• Post-merge: blender-smoke.yml has no push trigger — cite PR-head smoke, not merge SHA. Confirm release.yml fired and tagged (expect v0.47.0). If nothing fired, report it.

Evidence labels

Claims above are marked live-run-proven or inspection-only. Smoke matrix resolution and the 5.2 canary red run are not proven until CI on this PR.

[TMHSDigital]

5.2 canary — live-run-proven (then reverted)

Job (SHA 864fbf7): https://github.com/TMHSDigital/Blender-Developer-Tools/actions/runs/34401782403/job/102635164570

• Listing https://download.blender.org/release/Blender5.2/
• Downloading https://download.blender.org/release/Blender5.2/blender-5.2.1-linux-x64.tar.xz
• Print Blender version: Blender 5.2.1 LTS
• Canary step: ##[error]intentional 5.2 canary; revert after the red run — job conclusion failure
• Twin job Blender 4.5 smoke succeeded on the same run (https://github.com/TMHSDigital/Blender-Developer-Tools/actions/runs/34401782403/job/102635165199), so the new example already passed 4.5 CI before the canary came out.

Canary removed in 773e0e1. Waiting on 5.2+4.5 green + Socket.

[TMHSDigital]

5.2 smoke after canary revert — real failure, then fix

Post-canary SHA 773e0e1: 4.5 smoke success, 5.2 smoke failed at vse-cut-list --check-pixels (exit 12).

live-run-proven (.scratch/blender-5.2.1-windows-x64/blender.exe = Blender 5.2.1 LTS):

• TL sampled (0.949, 0.62, 0.102) (C amber) instead of crimson (0.85, 0.1, 0.22). Same numbers as CI.
• 4.5.11 and 5.1.2 pixel checks still pass with TL crimson.
• Cause: 5.2 COLOR strips bake readonly width/height from the scene render resolution at new_effect. CI creates at factory 1920 then renders 96×54; scale 0.36 of 1920 covers the tiny frame; C (ch 4) paints everything. 4.5/5.1 COLOR has no intrinsic size — scale is a fraction of the output frame.
• Probe: a lone COLOR strip with scale 0.36 does transform on 5.2; the bug is media size vs later output size.

Fix in 036cd19: set 96×54 before build_cut_list in check_pixels, and assert A.width, A.height == (96, 54) on 5.2. Local 5.2.1 pixels now: TL(0.851, 0.102, 0.22) TR(0.059, 0.749, 0.8) BL(0.949, 0.62, 0.102) BR(blend)(0.341, 0.349, 0.463) exit 0.

gn-modifier-inputs had already passed on 5.2 CI before this step.