fix(auto): preserve model-specific KV metadata in planning

[DivyamTalwar]

Summary

Fixes #25.

Auto transfers model fields then clears a.model, but omitted the existing KV-per-position fact. Nondefault presets and known atlas entries then used the generic fallback at nonzero context. Carry the existing bytes-per-position value through the args' KB boundary, preserving explicit overrides and unknown-model fallback.

Testing

Base: 252e5193902d466726da9af75047dfffff2ae662. Debian 12 Linux aarch64 in a nonroot disposable container, Python 3.11. The full existing smoke exits 0 with eight explicitly reported pre-existing optional research/hardware skips. It also prints baseline diagnostic skips where simulator/research dependencies are absent; no skipped check is counted as a pass. No Windows run or GPU/real-model benchmark is claimed.

The same final test files fail against unchanged production; the corrected branch gives:

Focused: 12 passed in 0.06s
Full applicable suite: 8 SKIPPED (a skip is not a pass):
all green

python tests/smoke.py
ruff check quantprobe
ruff format --check quantprobe
# Bandit medium-severity checks on changed package modules
git diff --check

All applicable commands above exited 0 locally. The snapshot is a complete upstream checkout plus this branch's exact changed-file bytes; no test is a copied production-function reimplementation. External I/O is mocked where stated. Dependencies and lockfiles are unchanged.

Additional repository audit: pre-existing failure

python audit.py exits 1 on both the unchanged production baseline and this candidate, with byte-identical audit output. It reports five existing scored preregistrations missing their Wired into statements. These findings were neither altered nor silenced. Package smoke, focused regressions, Ruff, and Bandit checks above pass; the repository audit does not pass. Full verification and maintainer acceptance therefore remain outstanding.

• MISSING 2026-08-09-quant-recipe-vs-naive-benchmarks.md: scored but never says where the finding went. Add a 'Wired into:' line naming the code path and test, or 'Wired into: nothing - ' if it correctly changed nothing.
• MISSING 2026-08-11-4b-ceiling-chain.md: scored but never says where the finding went. Add a 'Wired into:' line naming the code path and test, or 'Wired into: nothing - ' if it correctly changed nothing.
• MISSING 2026-08-11-format-vs-capability-decomposition.md: scored but never says where the finding went. Add a 'Wired into:' line naming the code path and test, or 'Wired into: nothing - ' if it correctly changed nothing.
• MISSING 2026-08-14-qwen38-27b-hybrid-fragility.md: scored but never says where the finding went. Add a 'Wired into:' line naming the code path and test, or 'Wired into: nothing - ' if it correctly changed nothing.
• MISSING 2026-08-16-rank-predicts-fragility.md: scored but never says where the finding went. Add a 'Wired into:' line naming the code path and test, or 'Wired into: nothing - ' if it correctly changed nothing.

Security And Data Access

No credential, production-data, authentication, read-only guardrail or privileged workflow changes are included. Tests use synthetic inputs and disposable paths. No new benchmark, fitted law, or hardware capability is claimed.

Notes

No numerical law, fitted constant, hardware measurement or performance result is added. Before standard/custom runtime handoff, restore the original explicit KV flag so the actual downloaded GGUF header can override pre-download metadata. Tests use existing rows and captured real argument flow with external I/O mocked.

AI-assisted implementation, isolated same-provider source review, and controller regression checks are disclosed. They are not maintainer approval, cross-vendor certification, or hosted CI. One focused, signed-off commit; no generated logs, personal config, model weights or worktree state is included.

Draft pending upstream CI and maintainer review. Companion changes touching the same module/test hook may require rebasing as they land; no combined branch is being submitted.