revamp

Large rewrites preserve behavior by separating mechanical translation from redesign. Source: https://bun.com/blog/bun-in-rust. Any executor follows repository model routing; invocation alone authorizes no agents.

1. Baselines before any code

• Freeze known defects so regressions remain distinguishable from old bugs.
• Benchmark throughput, memory, artifact size, and latency on the same workload/hardware.
• Count assertions; exit with none skipped, weakened, or deleted.
• Freeze feature-fidelity assertions; preserve intended branding, motion, interaction and data semantics.

2. Tests are implementation-independent

Tests target the language-independent CLI/API, not the replaced implementation; decouple first if needed. Add held-out representative consumers/projects. Known corpus parity proves that corpus, not complete compatibility.

3. Mechanical first, idiomatic later

Translate 1:1: preserve architecture/names/structure; refactor after parity. Mixing port/redesign obscures failures. Record big-bang when bridges would linger, incremental when shipping must continue.

4. Trial run before fleet

Port three representative files through translate -> compile -> adversarial review -> tests. Use the trial to calibrate prompts, checks, and queue before scaling.

5. Mechanical work queue

Rank compile errors. Without delegation, work sequentially under one owner. Authorized delegates use isolated worktrees, commit only owned scope, and never stash/reset/edit another.

6. Adversarial review with fresh context

Review from diff and contract, assuming semantic drift. At a non-trivial PR or ship endpoint, use only repository-permitted foreground cross-model review; otherwise perform a fresh inline evidence pass.

7. Semantic-equivalence traps

Probe release-erased assertions/side effects, integer overflow/bounds, copy-versus-reference defaults, recursion/stack limits with deeply nested fixtures, and locale/encoding defaults. Every discovered trap becomes a fixture.

8. Repair the process

When an executor fails, fix the prompt/checklist/queue and rerun rather than hand-fixing output. Humans inspect validation, spot-check, ensure CI kept tests, and own merge.

Exit

• All prior tests pass; zero skipped/deleted; assertion count reconciled.
• Same-workload/hardware benchmark meets the recorded baseline.
• Bug ledger distinguishes pre-existing defects and new regressions.
• Every diff receives adversarial review under repository policy.
• Rollback is documented; old implementation stays shippable until exit.