Close the arbitrary-clean-workspace gap for two bit targets

[petalas]

Summary

arbitrary finite clean workspace
  remove inactive image coloops in a shortest word
  restrict five calls to at most four total scalar roles
  exhaust 69,728 reduced words × 117 address collision patterns
  find zero two-target candidates

Proves the exact six-call minimum for two invertible rational 3-by-3 targets in the formal abelian bit model, with any finite number of zero-initialized, zero-restored auxiliaries. Rank-one idempotent calls and the stated free scalar mixing are included; free address maps are not.

Adds only research/bit-clean-workspace-compression/, with proof, separate C++ census/Python oracle, controls, exact sources and registered replay. Continues waynenilsen's PR2167 at cea5d7be431b09877848354f04d5f2ba512339fd and its canonical clean-ancilla handoff. That predecessor's one/two-auxiliary result and Ali Romman/PR468, eumemic/PR288 and huxint/PR203 retain credit. No predecessor engine or recorded census is imported. The different whole-endpoint requirement in PR2900 is explicitly distinguished.

Evidence

• Before: PR2167 explicitly leaves arbitrary clean-auxiliary counts open.
  After: the written coloop/workspace reductions and complete reduced census close that formal-model gap. All 8,158,176 word/pattern cases reject two rank-three targets. Six-call attainment is checked coefficientwise with distinct integer/fractional targets.
• The independent Python vector oracle agrees on every dimension-two word and deterministic higher-dimensional probes: 1,223 words and 143,091 relation cases. Six real CLI test groups pass, including normal/optimized Git-free relocation and actual omitted-quotient, skipped-census and strengthened-threshold implementation mutants that must fail.
• Registered CPython 3.12.15 replay and optimized CPython 3.13.5 replay pass on Linux x86_64. GCC 14.2 and Clang 19.1 produce identical census output. ASan/UBSan output agrees byte-for-byte. All 144 research-workflow tests, scoped Ruff checks and committed-byte preflight pass locally.

python3.12 -B scripts/research_package.py --package research/bit-clean-workspace-compression

Started at canonical main 0706e744cbda1e4d0da5a6c569ad9b47876c6960, priority revision 5. Refreshed through 475fec190d5ad746bd4690574dca30d6901acdaf; relevant guidance, selected record and live scientific inputs are unchanged. Final complete overlap covers 492 open PRs, 2,115 closed PRs and 1,585 Discussion records. Metadata coverage is not a novelty guarantee. No advance reservation was posted.

Limits: dimensions three/four have one full census implementation with partial independent oracle coverage, not two full implementations. No proof-kernel or independent human review, original predecessor replay, supplier-native compilation, hosting/precision/all-size admission or complete multiplication theorem is claimed. Hosted CI is separate from these local checks. No new kappa or selected-frontier change.

Merge danger

Door: two-way. One additive, opt-in research package.

Blast radius: research. Frozen predecessors, shared tools and the selected frontier remain unchanged.