Skip to content

Commit 2fa58df

Browse files
committed
Consolidation: overlap receipt for the fourth100 witness-MANIFEST pin family (#636/#638/#639/#640/#642/#644/#646)
1 parent 1f7a934 commit 2fa58df

18 files changed

Lines changed: 42314 additions & 0 deletions

‎notes/task-card-agent-22.md‎

Lines changed: 74 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,74 @@
1+
# Task card (autoresearch.md section 3)
2+
3+
Task ID / proposed scope key: `overlap-receipt` (consolidation track), scope key
4+
757723e3fd38f20fa774, refined with nonce 6bc45fb7886eb99e0494a94781da0e58.
5+
6+
Base commit; exact predecessor and dependency heads: base main
7+
1f7a934975e5eacaf83511c9ac870263bf095e05. Predecessor evidence:
8+
research/community-reconciliation-fourth100-audit/{ARCHIVE.json,REVIEW.json},
9+
docs/research/community-reconciliation-fourth100-review.md,
10+
docs/research/pr-ledger.md (+ pr-ledger.json), CONTRIBUTORS.md. No code
11+
dependency on any open PR head.
12+
13+
Concrete object and question: the seven PRs of the fourth100 reconciliation
14+
whose contributed-package replay fails at the same vendored-witness MANIFEST
15+
pin assertion (`assert sha(root/'MANIFEST.json') == src[...]` in
16+
verify.py line 39): #636, #638, #639, #640, #642, #644 (vectorsss / Chi Zhao)
17+
and #646 (jeremydiviney). For each exact recorded head: (a) is the recorded
18+
replay failure a content defect of the vendored witness package (registered
19+
MANIFEST hash vs actual blob) or a checkout/normalization artifact, and (b)
20+
what exact content of the PR set is already byte-retained in main's tree or in
21+
pinned archives, with complete predecessor credit?
22+
23+
Scope status: refined. The card's question ("what content and credit of a small
24+
named PR set has already been retained") is kept; the concrete set is chosen
25+
because the ledger records one shared, precisely described unresolved question
26+
for it and the fourth100 receipt is the existing receipt to audit.
27+
28+
Existing restriction being changed (or exact claim being audited): none
29+
changed. Auditing the fourth100 disposition receipt's unresolved question for
30+
this family ("Required contributed-package replay has not completed
31+
successfully").
32+
33+
Method; explicit parameter ranges/candidate alphabet: exact git-blob-identity
34+
comparison. Per PR head (git tree via the GitHub git/trees API at the recorded
35+
head SHA) every changed path's blob SHA is matched against local main's tree
36+
(git ls-tree -r 1f7a9349) and against pinned archive inventories; the witness
37+
MANIFEST pin is recomputed from the exact bytes returned by the contents API at
38+
the same head. No heuristics, no floating point; the domain is exactly the
39+
seven recorded heads and their changed-path sets.
40+
41+
Nearest overlapping PRs and why this adds something: PR1572 (receipt for
42+
PR148's 14 unresolved paths) is a different object; the live claim in
43+
discussion 1166#discussioncomment-18915531 covers the selected-replay tooling
44+
family #1101/#1142/#1151/#1277/#1361/#1509/#1646; the PR99->PR103 retention
45+
claim (discussion 1166#discussioncomment-18912011) is another object. A
46+
find_overlap pass with the refined object terms (witness635, "combined cold
47+
ends", vectorsss, pr643-improvement) returned only the PRs themselves and
48+
their downstream credit mentions; no receipt or audit of this family is in
49+
progress.
50+
51+
Optimistic full cost budget; required observable improvement: no kappa claim
52+
and no moment pricing apply (documentation/consolidation). Required
53+
improvement: every one of the seven unresolved ledger questions is replaced by
54+
an exact, checkable status (complete integration / partial retention / changed
55+
head / content defect), and any missing predecessor attribution is named.
56+
57+
Success evidence; falsification test; time/compute cap: offline stdlib checker
58+
recomputing the per-PR blob-retention counts and the MANIFEST pin verdicts from
59+
committed JSON inputs, with omission/substitution rejection controls; the
60+
smallest falsifying instance is one PR head (#636) where the recomputed pin
61+
disagrees with the recorded failure. Compute cap: single-digit minutes of API
62+
fetches plus one local pass; session timebox keeps the final quarter for
63+
verification and handoff.
64+
65+
Predecessor authors; intended artifact and reproduction command: vectorsss /
66+
Chi Zhao (#636/#638/#639/#640/#642/#644), jeremydiviney (#646), the fourth100
67+
reconciliation authors and the ledger maintainers. Artifact:
68+
research/pr636-family-witness-manifest-receipt/ (README with Open next
69+
questions, receipt JSON, offline checker). Reproduce:
70+
`python3 -B scripts/research_package.py --package
71+
research/pr636-family-witness-manifest-receipt` (and its --check-only
72+
validation before submission).
73+
74+
Attribution: Aaron Wang / navinor, with help from opencode.
Lines changed: 172 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,172 @@
1+
# Overlap receipt for the fourth100 witness-MANIFEST pin family
2+
3+
Scope key `757723e3fd38f20fa774` (`overlap-receipt`, consolidation track,
4+
priority revision 3, chooser nonce `6bc45fb7886eb99e0494a94781da0e58`),
5+
**refined** to one small named PR set: the seven entries of the fourth100
6+
reconciliation whose contributed-package replay fails at `verify.py` line 39
7+
with `AssertionError: larger_gain manifest pin` —
8+
[#636](https://github.com/CrocSwap/integer-mult-bounds/pull/636),
9+
[#638](https://github.com/CrocSwap/integer-mult-bounds/pull/638),
10+
[#639](https://github.com/CrocSwap/integer-mult-bounds/pull/639),
11+
[#640](https://github.com/CrocSwap/integer-mult-bounds/pull/640),
12+
[#642](https://github.com/CrocSwap/integer-mult-bounds/pull/642),
13+
[#644](https://github.com/CrocSwap/integer-mult-bounds/pull/644)
14+
(vectorsss / Chi Zhao) and
15+
[#646](https://github.com/CrocSwap/integer-mult-bounds/pull/646)
16+
(jeremydiviney). Base main `1f7a934975e5eacaf83511c9ac870263bf095e05`.
17+
18+
This is a source-bound disposition receipt: a finite certificate of git-blob
19+
identities and one diagnosed source-pin defect. It is **not** a kappa claim, a
20+
construction, or a mathematical result, and it changes no disposition: the
21+
seven PRs stay open, the selected frontier is unchanged, and no closure is
22+
requested.
23+
24+
## What was audited
25+
26+
The existing receipt under audit is the fourth100 reconciliation
27+
(`research/community-reconciliation-fourth100-audit/`,
28+
`docs/research/community-reconciliation-fourth100-review.md`) as recorded in
29+
`docs/research/pr-ledger.json`, which leaves all seven with the identical
30+
unresolved question "Required contributed-package replay has not completed
31+
successfully" and identical evidence text
32+
`AssertionError: larger_gain manifest pin`. For each recorded head I checked
33+
(a) the source-pin block of the package's own `verify.py` (lines 36-42: package
34+
MANIFEST tree, each pinned source's `MANIFEST.json` pin, and each pinned file)
35+
against the exact bytes committed at that head, and (b) the byte-retention of
36+
every changed path against main's tree at the base commit. Method: exact
37+
git-blob identity only (SHA256 of raw bytes for pins, blob SHAs for retention);
38+
no heuristic matching. The fetching side is `tools/audit_witness_pins.py`, the
39+
receipt builder is `tools/build_receipt.py`.
40+
41+
## Findings
42+
43+
**F1 — the seven failures share one cause: a drifted `larger_gain` pin.** All
44+
seven SPECs register `larger_gain` with
45+
`manifest_sha256 = 54d67f45ad520f4a134e069336bd0ff639aaeada28d13539298e4f2655c9692d`,
46+
which is exactly `research/larger-gain/MANIFEST.json` at the pinned PR569
47+
revision `897213a855b7fb8a1b8b0e552a2f5c42e6c1b326`. The same path **as staged
48+
into main** (commit `df417acd`, "Stage frozen PR family through 588") hashes
49+
`73c186c2807c703e43293c7cf785557a046958c245fa812e433f7fe208cdef6d`, and this
50+
drifted copy is what every one of the seven recorded heads carries in its base
51+
tree (checked at each head). The two manifests differ in exactly two `files`
52+
entries: `NOTICE` and `README.md` (both 64-hex digests changed; byte length
53+
40,559 in both). The staged package is self-consistent at its own revision:
54+
main's `NOTICE` and `README.md` hash to the staged manifest's values, not to
55+
PR569's. The other five pinned sources (`multicold`, `structural_gain`,
56+
`joint_prefix`, `frame_coalescing`, `direct_delivery`) still match their pins in
57+
the live repository packages.
58+
59+
**F2 — the six vectorsss packages vendor the pinned revision byte-exactly.**
60+
For #636/#638/#639/#640/#642/#644, each source pin resolves to the package's
61+
own `vendor/<name>/MANIFEST.json`, and every one matches its registered
62+
`manifest_sha256` in raw bytes (zero CRLF bytes anywhere, so the repository's
63+
text normalization is not involved). All their pinned files match too. Run
64+
through their own `verify.sh` source overrides, these six pass the pin block
65+
that the reconciliation recorded as failing; they evidently ran with the
66+
sources resolved to live repository packages (as #646's defaults do), where
67+
`larger_gain` is the first pin that fails. #646 has no vendored copy at all:
68+
its `verify.py` defaults resolve `larger_gain` to the live
69+
`research/larger-gain`, which reproduces its failure on any checkout of main.
70+
71+
**F3 — retention is partial, and dominated by re-vendored predecessor
72+
content.** Against main's tree at the base commit (exact blob identity):
73+
74+
| PR | changed paths | byte-retained in main (other path) | not byte-retained |
75+
|---|---:|---:|---:|
76+
| #636 | 348 | 285 | 63 |
77+
| #638 | 347 | 281 | 66 |
78+
| #639 | 340 | 281 | 59 |
79+
| #640 | 339 | 282 | 57 |
80+
| #642 | 354 | 281 | 73 |
81+
| #644 | 355 | 280 | 75 |
82+
| #646 | 186 | 77 | 109 |
83+
84+
No path matches at the same path, and no path exists in main with different
85+
content, so retention is entirely "same bytes at a different path": the
86+
vendored copies of the six pinned source packages already live in main under
87+
their own `research/` directories. The unretained remainder is each package's
88+
own witness, tools, data and result files (listed path by path in
89+
`evidence/pr{N}-audit.json`).
90+
91+
**F4 — the six vectorsss members share their vendored core with each other**
92+
(250-268 shared blobs pairwise; #646 shares 31-33 with each). A reviewer can
93+
treat the six as one content family plus six small per-PR deltas.
94+
95+
**F5 — predecessor credit is present.** `CONTRIBUTORS.md` credits vectorsss for
96+
#636/#638/#639/#640/#642/#644 (line 835) and jeremydiviney for #646 (line
97+
833). No missing attribution was found for this family; the downstream
98+
incorporated work that credits vectorsss's owner-search methods (e.g. #654,
99+
#671, #676, #699) is separate and unaffected.
100+
101+
## Reproduction
102+
103+
```sh
104+
python3 -B research/pr636-family-witness-manifest-receipt/verify.py
105+
python3 -B scripts/research_package.py --package research/pr636-family-witness-manifest-receipt
106+
```
107+
108+
`verify.py` is offline and standard-library only. It recomputes the drift
109+
identity from the two committed MANIFEST copies (and against the working tree's
110+
`research/larger-gain/MANIFEST.json`), re-derives every pin verdict and every
111+
retention class from the per-PR audit rows against `git ls-tree -r <base>`,
112+
recomputes the cross-PR shared-blob counts, runs a positive fixture of the
113+
classification rule, and runs four failure controls (tampered digest,
114+
substituted manifest diff, omitted row, substituted class) that must each be
115+
rejected. The evidence copies under `evidence/` are byte-exact unmodified
116+
copies of two `MANIFEST.json` revisions, kept so the digests are checkable
117+
without network access.
118+
119+
## Limitations (not checked)
120+
121+
- Head-side blob identities rest on the GitHub pulls/contents APIs at the
122+
recorded head SHAs; the seven PR trees are not fetched into the reviewer's
123+
git, so `verify.py` re-derives the main-tree side only.
124+
- The `multicold` sources declare `full_tree: true`; their whole-tree manifests
125+
were not recomputed file by file (the recorded failure is earlier, at the
126+
`MANIFEST.json` pin).
127+
- The seven packages' full contributed replays (construction, owner search,
128+
admission) were not executed. This receipt resolves the recorded pin failure
129+
and the retention/credit question only.
130+
- That the fourth100 harness resolved sources to live packages is inferred
131+
from the uniform `larger_gain manifest pin` assertion plus the pin data; the
132+
harness's exact invocation was not recorded in the ledger.
133+
- Retention matching is byte identity. Content rewritten equivalently would be
134+
reported as not retained; no such case appeared at these paths.
135+
136+
## Open next questions
137+
138+
1. **What did the fourth100 harness actually pass to `verify.py`?** It matters
139+
because the six vectorsss packages pass their pin block via their own
140+
`verify.sh` overrides, so their ledger questions may be resolvable by a
141+
re-run rather than a content fix. First step: re-run #636's `verify.sh` pin
142+
block (`--check-pins`) in a byte-preserving checkout of its recorded head
143+
`679621828a85bac68af869fb0e55b54508e6b14b`. Blocked by: the recorded
144+
traceback stops at the assertion and does not record the harness's source
145+
resolution.
146+
2. **Should the `larger_gain` pin drift be reconciled in main?** The staged
147+
`research/larger-gain` package is self-consistent but is a later revision
148+
(NOTICE/README.md changed) than the pinned PR569 revision that all seven
149+
packages register; every live-package resolution fails until someone either
150+
re-pins or vendors the pinned revision. First step: confirm with the
151+
maintainers whether `df417acd`'s NOTICE/README.md rewrite was intentional
152+
(it looks like license/attribution staging) and then update the seven SPECs
153+
or restore the pinned revision. Blocked by: maintainer decision on pin
154+
policy; this receipt does not change any pin.
155+
3. **Is any of the unretained content worth retaining for reuse?** The 57-109
156+
unretained paths per PR are mostly per-package witness outputs; a bounded
157+
follow-up could classify them into method code worth retaining versus
158+
generated evidence. First step: classify the `not_retained` lists in
159+
`evidence/pr{N}-audit.json` by extension and directory. Blocked by: nothing.
160+
161+
## Credit and attribution
162+
163+
Predecessor authors: vectorsss / Chi Zhao (#636/#638/#639/#640/#642/#644) and
164+
jeremydiviney / Jeremy Diviney (#646); the fourth100 reconciliation authors
165+
for the audited receipt and recorded evidence; the PR569 (`larger_gain`)
166+
authors and the `df417acd` staging authors for the two manifest revisions
167+
compared here. The receipt follows the source-retention receipt style of
168+
[#1572](https://github.com/CrocSwap/integer-mult-bounds/pull/1572) (PR148),
169+
which audited a different family.
170+
171+
Prepared by Aaron Wang / navinor with help from opencode. Substantial AI
172+
assistance is disclosed; no independent human review is claimed. Apache-2.0.

0 commit comments

Comments
 (0)