[kvlonge]

Why

Keep the config builder's native result instead of throwing it away and converting Python dictionaries back into nodes during planning.

Changes

• Retain the built tree across fresh, inherited and reused configs and final model-record replacement.
• Strip header metadata and replace expanded hooks in the retained tree at the existing handoff.
• Preserve public Python records, custom-key handling and lazy opaque containers. No new boundary exceptions or budget increases.
• Extend the real DuckDB CLI planning regression: ordinary and unused-extra config cases now require zero reconversions (baseline: 132).

Verification

• 26 focused integration tests and exact make check-ci passed after installing all extras.
• env --default-signal=INT UV_NO_SYNC=1 make test: 12,197 passed. The tool environment otherwise inherits ignored SIGINT; all seven cancellation cases pass with normal signal handling.
• One bounded independent review found no blockers; duplicate analysis and private hygiene scans passed.
• Dense/fresh 3,000-model comparisons on the same eight-core orb: three alternating pairs for uncached, cache-writing, warm and edited compiles; all 24 report/artifact identities and diagnostic counts match. Phase timings retained. Cold wall/CPU ratios remain below 1.05; all RSS ratios below 1.04. Warm differences are at most 0.04s CPU.
• Both 10,000-model guards passed locally under CI's exact 4 GiB cap. The baseline also completed without OOM but failed its edit-time guard. After preserving the initial hosted OOM, one diagnostic rerun passed both guards in 310.15s; no budgets changed. All required CI checks passed.
• No external real project was available; no end-to-end speedup claim. Rebased onto the merged predecessor with identical source trees.
• No DuckShop addition: internal ownership refactor, not a new project shape or behavioral change.