feat(research): price escape from refined P104 intervals

[petalas]

Summary

Extend PR #2136's survivor separator to a quantitative obstruction for added paid rank, while allowing arbitrary nested subdivision of all retained P104 transport obligations.

56 independent R-side source arrays
8 original entering array positions + at most B added positions
B >= 48
retained no-credit mass >= 60,840 / 60,736 = 585/584 > 1

Only research/p104-refinement-rank-budget/ is added. It contains the scoped proof, source-bound verifier, complete receipt, nine tests and an opt-in replay hook. The scientific base is canonical main 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, priority revision 4. Canonical claim and status.

Evidence

• Before: PR Rule out identity from nested P104 transport refinements #2136 proves the unaugmented 56-array bound but does not price escape. PR Prove P104 transport floor with arbitrary added storage #2127's 96-transport result retains unsplit occurrences.
• After: the written transcript argument gives B >= 48 with arbitrary nested refinements, added transports and source-independent storage. The exact no-progress-credit screen from PR Prove sharp P104 temporal rank bound and correction-budget obstruction #2082 then fails at zero saving. A changed coverage proof must credit more than 104/9 ranks per call even at the minimum added cost.
• Registered CPython 3.12.15 replay passes normally and with -O; all five semantic rejection controls and nine CLI/export tests pass. The nine tests also pass under -O. Separately, all 65 repository research workflow tests pass. A separate CPython 3.13.5 optimized replay agrees. Environment: Linux x86_64, standard library only.

python3.12 scripts/research_package.py --package research/p104-refinement-rank-budget
python3.12 -I -O -X dev -W error -B research/p104-refinement-rank-budget/test_verify.py
python3.12 -B -m unittest discover -s tests -p 'test_research_*.py'

The new contribution is the added-rank implication and budget decision, not a new separator. Credit Tairitsu-Aya/PR #2136, Ross Richardson/PR #2127, David Arias/PR #2082 and the earlier cuts; Rohan Arun/PR #1972, Chafik Boukhalfa/PR #1991 and their geometry/native/coverage predecessors remain credited in the README. Exact heads and source hashes are recorded.

Merge Danger

Door: two-way.

Blast Radius: research.

This is an author-checked scoped proof and finite premise check, not a supplier or exponent improvement. It retains original obligations without removal, fusion or branching. Its moment conclusion assumes the explicit no-progress-credit coverage contract. Changed producers, replacement paths and justified coverage credit remain open; Discussion #2131 and PR #2139 are not excluded.

Native phases, the original scalar endpoint, chronological hosting, precision/tape, all-size admission and the full selected cold replay were not rerun. The transcript proof is not formal-kernel verified or independently human-reviewed. The selected record and frozen evidence are unchanged. Live overlap was refreshed before submission; exact coverage and its limits are in TASK.json.