docs: README and docs/models.md describe 0.3.3

[midagedev]

README and docs/models.md describe 0.3.3. Docs only.

What

• README status is 0.3.3:
  • MiniCPM-V 4.7 images (gpu-vision, vision, serve: MiniCPM-V 4.7 images end to end, a tile capped at 70 tokens #147, --mmproj on the qwen3 seat, a tile capped at 70 tokens);
  • the two Qwen3.8 limits 0.3.2 listed, fixed in gpu-gates: a reduced one-slot qwen38 draft stays up, and asked_draft names the ask before the yield #121 and gpu-gates: (pg) plans card experts on a machine that counts their route scratch, and reads the drafted plan's room from MtpPlan #255;
  • the release pass points to the 0.3.3 release notes.
  • The 0.3.2 release-pass table is gone; the release notes carry 0.3.3's.
• The image input row names MiniCPM-V 4.7.
• Known issues (0.3.3):
  • the two 0.3.2 lines fixed by gpu-gates: a reduced one-slot qwen38 draft stays up, and asked_draft names the ask before the yield #121 and gpu-gates: (pg) plans card experts on a machine that counts their route scratch, and reads the drafted plan's room from MtpPlan #255 are gone;
  • the Qwen3.6 real-tier e2e split clause's run-to-run distance is added: worst 4.879e-1 at call 34 of 82 against FREE_BAND 0.26 (gate_qwen35moe_e2e.rs:914), RTX 3090, the engine's gpu, gpu-gates, model, refset: Nemotron 3 Nano 4B on the card and bloomery-serve --model nemotron #259 tree (box tree 770ed45a), 2026-10-15 18:33 KST (specs/engine/st-batch11/g-gate-gpu-qwen35moe-e2e.log:857); reruns at 1.6–1.8e-1 (perf's note).
  • gate-gpu-qwen38-serve's real tier 70/72 stays as 0.3.2's measured line (f382b474), marked not re-run; the release pass re-checks it.
• The tarball names and the release link are 0.3.3's.
• docs/models.md: MiniCPM-V 4.7 takes images with --mmproj since 0.3.3.
• The Nemotron 3 Nano 4B rows and its prompt-speed line are a separate branch (docs033-nemo) that lands after gpu, gpu-gates, model, refset: Nemotron 3 Nano 4B on the card and bloomery-serve --model nemotron #259.

Evidence

Docs only; no code, gate or recipe changed.

Merge order

• Merge after the v0.3.3 tag exists: the status paragraph and the tarball line link releases/tag/v0.3.3.

Follow-up (not in this PR)

• The README's split line names its card but not its commit: the numbers come from the engine's gpu, gpu-gates, model, refset: Nemotron 3 Nano 4B on the card and bloomery-serve --model nemotron #259 tree before it merged. The 0.3.3 release pass replaces them with a run at the tag, with its commit.

Pre-review CLEAN at 38f7e8c.

🤖 Generated with Claude Code

[github-actions]

Verdict: CLEAN: No must-fix finding remains in this documentation-only diff.

Non-blocking

⚠ Attach the split-clause measurement evidence: README.md:64 gives new distances without linked run output.
The body identifies the first run's tree but supplies no driver line, model file or gate command/output for either run.
Attach that evidence and name each run's commit beside the figures. This keeps later tag results distinguishable from the pre-merge measurements.

Checked

Duplicates: 0 new functions, types or constants; no dryer pairs to read.
Complexity: no changed Rust functions; no crapper candidates.
Silent failure: 0 changed sites; oversized MiniCPM tiles return a named shape error.
Tests: none added or changed; mutation checks do not apply. Runtime tests were not run for this documentation-only diff.
Gates: no thresholds, clauses, tests or pins changed.
Escape catalog: no C1–C4 patterns matched.

Notes

Dryer and crapper did not run because no Rust file changed.

_{🔨 Shingler, bloomery's review bot: approves a clean pull request and requests changes on a must-fix (how it works)}