hop-A inputs — the recon’s raw data, BANKED (was: stranded in a session scratchpad)

The corpus-hop audit found the inputs behind RECON-go12312-diff.md living only in the recon session’s scratchpad (“Raw TSVs beside this file for re-derivation” — beside the report, but the report moved to the repo and the TSVs did not). A hop input that exists only in a session scratchpad re-derives from zero if that session dies. These are the same five files, re-derived independently on a second machine and verified against every count the recon itself published — which makes this bank stronger than a copy would have been: a copy proves the files were saved; a re-derivation proves the recon’s numbers are reproducible.

The files

file	rows	what
commits.tsv	83	<sha>\t<date>\t<subject> for every commit in go1.23.1..go1.23.12
files-by-commit.tsv	215	<sha>\t<file> — per-commit changed-file list (the compare API caps its files array; per-commit enumeration is ground truth, same rule the recon used)
files-unique.txt	161	sorted unique changed files, repo-wide
src-files-classified.tsv	150	<class>\t<file> for the src/ subset; classes cmd / runtime-tree / stdlib
roster.txt	162	the validated-package roster as read from docs/ValidatedTestPackages.md at this bank’s commit

And the rehearsal’s own raws, banked 2026-08-24 — ../REHEARSAL-go12312.md promised these “beside this file” and they were in a session scratchpad, one purge from gone. They are copied here verbatim, not re-derived:

file	rows	what
census-raw.txt	152	the full stdout of migrate-gorelease.ps1 -To 1.23.12 in census mode — the run behind the rehearsal’s §1 arithmetic (20 sites / 8 files, UNCLASSIFIED: none)
h3-files.tsv	160	<status>\t<file> for the compare endpoint’s file list — the rehearsal’s independent H3 derivation, a strict subset of files-unique.txt by exactly one cmd/ test file
h3-compare-head.json	1	the compare endpoint’s headline object, {"commits":83,"files":160,"status":"ahead"} — the truncation the recon predicted, preserved as the evidence for it

Derivation (reproducible from a clean machine)

git clone --bare --shallow-exclude=go1.23.1 --branch go1.23.12 https://github.com/golang/go gogit
git -C gogit fetch --deepen=1 origin   # REQUIRED: the shallow boundary commit is grafted, and a
                                       # grafted commit's --name-only lists its ENTIRE TREE
                                       # (13,261 phantom files); one deepen restores its parent
git -C gogit log --format='%H%x09%ad%x09%s' --date=short go1.23.12            # commits.tsv
git -C gogit log --format='COMMIT%x09%H' --name-only -83 go1.23.12            # files-by-commit

Classification rule (matches the recon’s headline arithmetic): cmd = src/cmd/**; runtime-tree = src/runtime* including runtime/debug (the recon’s package-bucketed table lists runtime/debug under stdlib-visible, but its headline’s 42/59 split buckets it with the runtime tree — both conventions differ by exactly that one file, runtime/debug/mod.go); stdlib = the rest.

Verification against the recon’s published counts — all exact

reading	recon	this bank
commits in range	83	83
unique changed files	161	161
under src/	150	150
outside src/	11	11
src/cmd/**	49	49
runtime tree	42	42
stdlib-visible remainder	~59	59
roster rows	162	162

⚠ The shard map — still unbanked, and its reserved-set gap is WIDER than the audit recorded

CLOSED, both halves, 2026-08-24. The draft and its generator banked into this directory verbatim (shard-map-draft.md, shardmap.py, e0d8930e1) — as found, so the next commit’s subject was visible rather than smuggled — and the recommendation at the end of this section then landed: shardmap.py derives its reserved set from $longTimeouts at generation time (549b4e556, lines 71–88), so the copied list is gone rather than corrected. PLAN-hop-campaign.md §4.3’s static table is marked SUPERSEDED BY GENERATOR, and the derive-never-copy rule is generalized into GoCorpusMigration.md §3.2. The paragraph below is kept as written, because it is the reasoning that produced the fix.

The 28 KB shard-map draft and its generator (shardmap.py) remain on the coordinator machine’s session scratchpad — not reachable from this lane, so not banked here; that half of the audit item stays open until the coordinator pushes them (or the generator is rewritten against banked data).

What this lane could verify: the audit said the map’s reserved set is missing go/parser (90m) and crypto/internal/mlkem768 (30m). Measured against the LIVE $longTimeouts table (src/run-validated-sweep.ps1:495) the gap is now three — crypto/tls (30m) joined after the audit — and two floors moved under it (crypto/dsa 60m→120m, archive/zip 30m→60m). The table has grown twice since CLAUDE.md described it as four rows. Recommendation for the map’s banking: the generator should DERIVE its reserved set from $longTimeouts at generation time, never carry a copied list — a copied list has already drifted twice in the map’s short life, which is the same hoist-vs-derive defect this campaign fixed in _paths.ps1 the same week.