[petalas]

Summary

The local P104 fusions in #2662 and its newer #2712 successor preserve rank mass, but their positive moment gains do not yet quantify an allowance for additional hosting charges. This adds an independently implemented, source-bound certificate for that missing cost implication.

PR2586 literal word -> PR2662 -> PR2712
       independently recounted frame ranks and profiles
       exact rational gain / one-added-H1 comparison

At the selected complex primitive saving, the gains relative to #2586 are 0.5273081047… and 0.5409352129… rank-one child units. One uncompensated H1 outweighs either local gain when caller width, role normalization and all other charges are fixed. Certified G=1 brackets are [0.003182398581, 0.003182398582] and [0.003101949576, 0.003101949577].

Only research/p104-fusion-cost-headroom/ is added. It includes exact source pins, three compressed literal words, retained notices, a written enclosure/monotonicity proof, receipt, controls and registered replay. Credit commdata2338/#2662, ywz65/#2586/#2712 and Nick Petalas/#2550, with the earlier lineage retained. The canonical claim and handoff and concurrent conditional consumer screen distinguish the scopes.

Evidence

• Before: predecessor profile improvements were qualitative; extra child charges remained open.
  After: exact enclosures establish the allowance, and also show both local comparisons afford three H1 children at the separate planning benchmark a=1/99.
• python3.12 scripts/research_package.py --package research/p104-fusion-cost-headroom passes on CPython 3.12.14: normal/optimized primary, six mathematical controls and six oracle/artifact test groups. Tests cover source/receipt corruption, Git-free read-only replay, recompression, output refusal and resource limits. Ruff diagnostics pass.
• No predecessor Python is imported. Source words are authenticated against exact Git blobs and SHA256; the selected benchmark is bound through its selected record and assembly. Separate internal arithmetic review agrees; no independent human or formal proof review is claimed.

Merge Danger

Door: two-way. Isolated research package and opt-in bounded replay.

Blast Radius: research.

This is a fixed-profile local budget, with no offsetting credits. Actual idle savings, sharing, changed consumers and changed profiles need a new bill. Native correctness, ownership, precision, the other 102 forward correction rows, all-size admission and a complete favorable supplier moment remain open. No selected-frontier or kappa change; no supplier cold-replay claim.