[petalas]

Summary

PR2743 leaves count tapes outside its complete first-node payload reserve. This adds a source-bound written count-storage proof for Jose Manuel/jmbarrancoml's unchanged PR2697 caller at a53fdf4bcbe33c6732f0bec1badc8c023380f9d1.

participating cardinality tapes <= 98, rounded to 100
symbols per tape <= B + 2
count content <= 100 * (B + 2), excluding LEFT
B = max(1, bit_length(V), indices.bits)

The inventory includes borrowed arguments, Count.literal's simultaneous raw/returned tapes, product/add temporaries, marginal padding, countdowns, the original-offset peel and final merge. The fixed source callbacks and absent optional observer are explicit hypotheses. This is one reserve category; numerical/control/directory tapes, LEFT markers, physical placement, larger returns and total reserve admission remain open. There is no supplier or kappa change.

Only research/selected-node-count-storage/ is added. It contains the proof, exact receipt, 43-routine inventory, bounded source archive, per-file source pins, standalone checker, 11 focused tests and registered replay hook. All 45 source exports independently match the pinned commit; predecessor notices and proof are unchanged. Credit Jose Manuel's caller, Nick Petalas's PR2708/2743 payload work, starship-s's PR2731 residual work, ywz65's directory/primitive work and retained lineage. PR2473's countdown time and PR2590's directory setup time are complementary scopes.

Task card selected-stock-envelope, priority 5, is explicitly changed to the live Discussion2319 handoff. The original complex-track suggestions were checked before taking this current reserve component. Scoped source/lifetime and checker reviews found two issues that were fixed: arbitrary observers can mutate borrowed counts, and preloaded Python module names could bypass source imports.

Evidence

• Before: complete caller payload evidence excludes count tapes, and a one-tape budget rejects Count.literal's real constructor overlap. After: the written source inventory gives 92 owned roots plus 6 borrowed counts, with transient padding explicitly priced.
• The complete GL2 node at V=15,552 observes 119 occupied count symbols, at most 27 tapes and 15 symbols per tape across 119,351,262 count writes. All 5,712 child entries/returns complete; direct cell censuses, independent returned-bit checks, borrowed-input/head restoration and count cleanup pass. A separate literal n3/d2 peel at V=729 measures 66 symbols, 18 tapes and 11 symbols per tape.
• Registered CPython 3.12.15 replay, optimized padding/peel controls and all 11 focused tests pass locally. The repository research-workflow suite passes 70 tests with 3 skips. Controls cover constructor underbudget, product/countdown behavior, inventory exhaustion, source/import/archive faults, recompression and typed receipts. Two internal reviewers checked the source proof and45 exact exports; no independent human, formal-kernel or hosted-CI pass is claimed.
• Full selected GL64 payload execution, the selected supplier cold replay and total numerical stock admission are outside this component. The 100 cap is conservative rather than optimal. Successful execution is a premise; inherited exception cleanup is separate.

uv run --python 3.12 --no-project python -Werror -B scripts/research_package.py \
  --package research/selected-node-count-storage
uv run --python 3.12 --no-project python -Werror -B -m unittest discover \
  -s tests -p 'test_research_*.py'

Started from current canonical main 9ed84862fe4cb6eb67b2572eca332c92df803a1f. Submission freshness and live-overlap coverage are recorded in TASK.json. Metadata coverage is not proof of novelty. Scientific source heads remain frozen.

Merge Danger

Door: two-way

Additive research package; production suppliers, selected result and predecessor packages are unchanged.

Blast Radius: research