CONTRIBUTIONS.md

Upstream contribution record

Snapshot assembled at 10 November 2026 · 22:01 CST (UTC+8) for Afloat16. Display dates use Asia/Shanghai. The source snapshot retains each observation time. Data snapshot.

397 verified merged author PRs · 36 upstream repositories · 420 default-branch author commits across 37 accepted-contribution projects

Since the 10 November 16:44 snapshot, 2 more PRs merged, both in serengil/deepface. The 13 upstreams above 10,000 stars account for 302 verified merged PRs.

30 upstreams have more than 1,000 stars, covering 382 directly merged PRs. Afloat16 appears in the top 30 on 17 of those projects’ Contributors graphs.

Verification method

• GitHub’s complete public merged-author search returned 395 unique PRs. The complete closed-author search returned 569 records, including 397 with merge timestamps. Direct PR checks again confirm the two records omitted by the merged-search index, Geoopt #258 and #259. Each repository and PR number is counted once. The source snapshot records the search-page observation times and the final incremental check.
• Both additions, DeepFace #1702 and #1703, were checked for the Afloat16 PR author, merge timestamp, final SHA and target branch. Their final commits are reachable in the current master history. DeepFace now has 33 directly merged author PRs and 33 independently enumerated default-branch author commits.
• All 420 default-branch author commits across 37 accepted-contribution projects were independently enumerated again. Claude-Mem and Craft Anything required two pages each. This metric is separate from merged-PR totals and the graph’s contributor counts.
• 27 of the 30 homepage positions were rechecked from the official Contributors graph using All, Commits and the default branch, excluding merge commits. This refresh read 26 numbered Afloat16 entries and 1 complete top-100 lists without an entry. A dash represents a complete top-100 absence check. The remaining 3 entries retain their earlier complete observations and original timestamps after browser-reading timeouts, as marked in the tables. GitHub’s graph definition.
• The separate adoption inventory retains 15 verified maintainer integrations and cherry-picks with their original dates and attribution. The bounded follow-up in the four previously recorded adoption repositories inspected 11 other-author merged PRs and their final commit messages from 10 November 16:44:06 to 21:26:26 Asia/Shanghai, finding no new attribution or known-source-reference candidates. Haystack #13498 was closed without merging and is excluded from both merged and adoption totals. This was a scoped candidate check. Interval evidence.

Current snapshot · Previous snapshot · Previous full record · Author PR search

Selected merges by project

The same seven projects each have one row, with 17 representative merged PRs in total. Selection follows each project’s own merge cadence and weighs the behavior changed in the final code, technical depth and inspected regression evidence. Rows follow their latest selected merge. Each PR below retains its actual merge date.

Latest selected merge	Upstream project · ★	Representative engineering	PR / merged commit
2026-11-10	Claude-Mem 115,655 ★	Serialize concurrent generation retries and cancellations, refuse incompatible restore snapshots before replacing live data, and order replica-vector cleanup after committed deletions.	#5650 · e7978b0 #5649 · c141e54 #5647 · d58a11d
2026-11-10	Craft Anything 25,574 ★	Reject incomplete GNU tar metadata before installation writes, bound retained Node SEA metadata, and recover ELF dynamic imports and exports from partial section layouts.	#1402 · 2174cac #1377 · ee73141 #1367 · a5ea19c
2026-11-09	DeepFace 23,667 ★	Preserve exact S3 keys and release response streams, validate FTP endpoints and connection cleanup, and reuse decoded queries for one-shot image inputs.	#1686 · 00e4fae #1685 · e7734e1 #1701 · a2d66b2
2026-11-09	COLMAP 13,193 ★	Restore Python 3 image collection with bounded retries and response cleanup, and make translation-angle evaluation invariant to reconstruction scale.	#4977 · 9f0736e #4932 · 27ffad4
2026-11-06	REA 448,836 ★	Preserve UTF-8 at diagnostic cutoffs, resume complete NDJSON messages after interrupted iteration, and honor cancellation when bounded reads finish.	#8907 · 3eff806 #8909 · a3621e5 #8908 · b153ba8
2026-11-05	Sentence Transformers 19,275 ★	Exclude padded tokens before CNN convolution and compute RRF retrieval metrics without inserting unretrieved positives into observed rankings.	#4323 · 7da542f #4114 · 6994c94
2026-10-22	LeRobot 28,792 ★	Apply V-JEPA 2's pairwise RoPE correction to VLA-JEPA's training-only world predictor.	#4835 · f9e9d73

Claude-Mem

#5650 — Serialize operator retries and cancellations with concurrent generation-state transitions · merged 2026-11-10 05:44 · e7978b0

Re-reads the scoped generation job under the same PostgreSQL row lock used by generation completion, evaluates the transition and updates state within one transaction. Queue operations and audit publication occur after that transaction.

Committed PostgreSQL integration fixtures use isolated schemas, real row locks, gated concurrent requests and lock-wait inspection. They cover retry conflicts after processing or completion, duplicate retries, cancellation racing with completion and simultaneous cancellations.

The integration suite requires CLAUDE_MEM_TEST_POSTGRES_URL and skips when it is absent. No database-backed test was executed during this review. Queue I/O remains outside the row-lock transaction.

#5649 — Reject newer-schema restore snapshots before replacing the healthy live database · merged 2026-11-10 05:52 · c141e54

Checks the staged snapshot's PRAGMA user_version against LATEST_SCHEMA_VERSION in the SQLite file-job process before checkpointing or backing up the live database. The staged WAL participates in the compatibility check.

Committed SQLite fixtures verify rejection preserves the live and snapshot bytes, creates no pre-restore copy, leaves a failed pending restore recoverable by the worker and detects a newer schema stamp in the WAL. Older, current and unstamped snapshots remain accepted.

The preflight checks the schema version stamp and existing SQLite integrity path. This review inspected the fixtures and did not run the restore suite.

#5647 — Schedule replica-vector cleanup after committed remote deletions and earlier writes finish · merged 2026-11-10 05:45 · d58a11d

Captures the replica's local row ID before deleting a tombstoned origin record, dispatches cleanup only after the SQLite transaction commits and orders it after earlier writes for that row. Local and replicated deletions share a serialized cleanup lane.

Committed fixtures cover observations, summaries and prompts, assert the replica row ID and matching document type, delay an earlier write to test ordering, discard cleanup on rollback and verify a simulated vector-service failure preserves durable sync progress and permits later cleanup.

Vector cleanup is best effort. Failures are logged and durable synchronization continues, so a failed vector-service deletion can leave fragments until a later cleanup. Tests use an in-memory SQLite store and a simulated vector client, and were not rerun.

Craft Anything

#1402 — Reject incomplete GNU tar metadata before an engine installation can replace existing files · merged 2026-11-10 01:27 · 2174cac

Requires GNU long-name and long-link records to fit their complete padded extent, handles arithmetic overflow and reports a precise incomplete-listing reason. The existing installer consumes that failure before writing extracted members.

Merged parser fixtures cover missing metadata padding and complete records. Installer regressions exercise short padding and overflowing extents for both long-name and long-link records, assert GarbledOutput and verify the previously installed tool remains byte-for-byte intact. A valid ordinary-member fixture preserves the existing padding compatibility.

The production change is in the tar parser. Installer edits are regression tests for the existing preflight rejection path. This review inspected the committed fixtures and did not execute them.

#1377 — Bound retained Node SEA metadata while preserving valid later assets and complete input validation · merged 2026-11-09 23:34 · ee73141

Shares an 8 MiB decoded-text budget across code paths, asset keys and execution arguments, caps retained entry counts and accounts for UTF-8 replacement expansion. Oversized fields are omitted whole, and skipped metadata still has to fit the source blob.

Merged integration fixtures check that the main entry consumes an entry slot, empty arguments respect count limits, invalid UTF-8 spends the decoded-byte budget, oversized keys do not hide later valid assets, truncated skipped fields fail and ordinary metadata remains complete.

Bounds the retained listing metadata in the Node SEA reader. It does not establish a whole-process or source-blob memory limit. Fixtures were read without a fresh execution.

#1367 — Recover ELF dynamic imports and exports when surviving sections omit the dynamic symbol table · merged 2026-11-09 23:35 · a5ea19c

Triggers segment-based dynamic-symbol recovery when the object reader has no dynamic symbols, even if other sections survive. It preserves local symbols and deduplicates recovered symbols, exports and imports within the existing limits.

Merged synthetic ELF fixtures cover 32-bit and 64-bit layouts in both byte orders, retain a .text section and a local symbol, and require each recovered dynamic entry, import and export exactly once. A complete dynamic-section control checks that recovery does not duplicate symbols.

Recovery uses load-segment dynamic metadata available in the binary. The committed regression fixtures were inspected, without running the test suite.

DeepFace

#1686 — Preserve exact S3 object keys and close downloaded response bodies on successful and failed reads · merged 2026-11-09 23:29 · 00e4fae

Keeps S3 key delimiters and leading path characters intact through listing, joining and reads, then scopes response bodies with closing so read errors still release the stream.

Merged mock-backed fixtures round-trip listed keys containing reserved characters and extra leading slashes, check exact bucket/key arguments for read and write operations, and require body.close exactly once after both a successful read and a raised read error.

Uses mocked S3 clients and streaming bodies. It does not demonstrate a live S3 integration run.

#1685 — Validate FTP endpoint identity, preserve credential semantics and close failed or stale connections · merged 2026-11-09 23:10 · e7734e1

Decodes URI credentials while preserving literal environment credentials, handles bracketed IPv6 endpoints and validates the expected port. It closes failed or stale sockets, avoids exposing credentials in endpoint errors and propagates protocol failures beyond missing-file replies.

Merged mock-based fixtures check literal environment credentials and decoded URI credentials, IPv6 identity, wrong-port refusal without credential disclosure, failed-login retry after socket cleanup, stale-connection replacement and SIZE error propagation.

The regression fixtures simulate FTP clients and protocol responses. No live FTP server was exercised during this review.

#1701 — Reuse the decoded query image so face search can consume a one-shot input stream once · merged 2026-11-09 22:08 · a2d66b2

Passes the image already loaded by find into face extraction, preserving the query after an input stream has been consumed and avoiding the duplicate decode path.

The merged regression supplies a one-shot source with a read counter, mocks the surrounding model/gallery operations and verifies both the expected match and exactly one read of the original source.

The test isolates the recognition dataflow with mocked model and gallery dependencies. The same final commit also includes workflow and authentication annotation edits, which are outside this selected contribution summary.

COLMAP

#4977 — Restore Python 3 Flickr collection with bounded retry handling and response cleanup · merged 2026-11-09 02:33 · 9f0736e

Uses Python 3 URL requests and decodes HTML before parsing image links. Responses are closed and retry attempts terminate after repeated timeouts.

Tests use fake HTTP responses to verify a downloaded photo and response closure, a successful retry and exhaustion after repeated timeouts.

The change affects the image collection utility.

#4932 — Make translation-angle errors invariant to reconstruction scale · merged 2026-11-02 16:40 · 27ffad4

Normalizes finite nonzero vectors using their largest absolute component before their L2 norm, avoiding the small-norm floor and large-magnitude overflow.

Regressions scale vectors from 1e-200 to 1e200 and check angular invariants, zero vectors, input preservation and reconstruction-level similarity transforms.

The change affects benchmark translation-angle evaluation.

REA

#8907 — Keep complete UTF-8 code points in bounded diagnostic excerpts · merged 2026-11-06 10:13 · 3eff806

Trims only incomplete code-point fragments at retained head and tail boundaries while leaving known malformed bytes visible as diagnostic evidence.

Regressions cover every two-chunk split, bytewise appends, malformed prefixes and bounded subprocess stdout and stderr.

The fix applies to byte-bounded diagnostic excerpts.

#8909 — Resume already-buffered NDJSON frames after iteration stops · merged 2026-11-06 10:12 · a3621e5

Drains complete pending lines even when a later push is empty or contains only a partial following frame.

A generator is closed after its first yielded line and the remaining complete lines must drain on an empty push. Hopper stream fixtures check partial pushes after an invalid response.

The fix applies to decoded NDJSON buffering and response consumption.

#8908 — Honor cancellation before returning from a completed bounded read · merged 2026-11-06 10:11 · b153ba8

Checks the abort state immediately after each awaited read as well as before initial work, including empty and fixed-size reads.

Boundary tests abort while a filesystem read is in flight for empty, exact-fit and limit-crossing inputs and check that the caller-owned cancellation reason reaches the configured error conversion.

The fix applies to bounded file-byte reads.

Sentence Transformers

#4323 — Mask padded token embeddings before CNN convolution · merged 2026-11-05 20:45 · 7da542f

Zeros masked token embeddings before applying convolution so padding values cannot influence neighboring valid tokens and pooled sentence embeddings.

Tests combine left and right padding, large and NaN padding values and multiple kernel sizes, with an unmasked control.

The change affects the CNN module when an attention mask is supplied.

#4114 — Evaluate observed RRF rankings with all known positives as normalization · merged 2026-11-02 18:54 · 6994c94

Computes MRR, MAP and NDCG from retrieved documents alone while normalizing MAP and ideal DCG against all known positives. Example rerankers receive the actual retrieved candidate set.

Regressions cover missing positives, a top-100 cutoff, empty and singleton results, complementary retrievers, tied scores and a zero cutoff.

The correction changes evaluation semantics and documented example metrics.

LeRobot

#4835 — Apply V-JEPA 2's existing pairwise RoPE correction to VLA-JEPA's training-only world predictor · merged 2026-10-22 22:00 · f9e9d73

Replaces frequency-block repetition with repeat_interleave(2, dim=-1), matching the adjacent-coordinate rotation layout and retaining the position axis for a single frequency. The correction targets the RoPE geometry used by the world predictor; it is an application of prior upstream work.

The final upstream commit changes the production helper and compatibility documentation. No regression-test file was added.

Not an original RoPE algorithm. The final merged diff contains the production helper and documentation only, with no added regression test file. Parameter names/shapes remain load-compatible, but existing world-predictor outputs and training losses change; pretrained quality requires separate evaluation. The world predictor is training-only, so no closed-loop robot performance, inference-quality or full-suite/GPU validation improvement is claimed. Tests were not rerun during this review.

Previously selected work

Earlier selections and their technical reviews remain in the 9 November 20:33 full record and its previous selections.

Complete merged-PR index

The index includes every directly merged author PR. Repositories below the homepage threshold retain their original graph observation dates. Homepage graphs that could not be refreshed also retain their prior observations and are explicitly marked retained with the original time. The two commit columns record the graph’s displayed contributions and the independently enumerated default-branch author history.

Repository	Upstream ★	Graph position · checked	Graph commits	Default-branch author commits	Merged PRs	Verified merged PRs
morluto/rea	448,836	#56 2026-11-10	6	6	6	#8907, #8909, #8908, #4524, #4414, #3919
thedotmack/claude-mem	115,655	#4 2026-11-10	109	108	108	#5631, #5625, #5643, #5634, #5633, #5642, #5635, #5632, #5630, #5629, #5628, #5627, #5626, #5622, #5624, #5649, #5648, #5647, #5650, #5651, #5588, #5646, #5620, #5586, #5611, #5587, #5580, #5579, #5645, #5640, #5582, #5623, #5609, #5608, #5589, #5585, #5621, #5583, #5641, #5638, #5636, #5610, #5607, #5606, #5644, #5637, #5619, #5617, #5618, #5616, #5615, #5614, #5613, #5612, #5602, #5639, #5605, #5603, #5604, #5533, #5534, #5561, #5560, #5559, #5558, #5545, #5557, #5556, #5550, #5549, #5548, #5547, #5546, #5572, #5571, #5570, #5569, #5568, #5567, #5554, #5529, #5566, #5555, #5528, #5562, #5564, #5565, #5563, #5552, #5553, #5551, #5574, #5573, #5532, #5536, #5539, #5538, #5531, #5524, #5523, #5537, #5526, #5525, #5462, #5458, #5367, #5340, #5341
huggingface/lerobot	28,792	— 2026-11-10	—	1	1	#4835
deepset-ai/haystack	26,982	— 2026-11-10 16:37 · retained	—	1	1	#13076
storytold/craft-anything	25,574	#2 2026-11-10	131	120	120	#1402, #1364, #1394, #1375, #1382, #1393, #1381, #1374, #1372, #1371, #1363, #1387, #1379, #1373, #1392, #1391, #1390, #1389, #1388, #1386, #1385, #1384, #1380, #1367, #1366, #1365, #1378, #1377, #1370, #1368, #1362, #1329, #1318, #1348, #1343, #1342, #1341, #1340, #1339, #1338, #1337, #1336, #1335, #1334, #1333, #1330, #1332, #1328, #1327, #1326, #1325, #1324, #1323, #1322, #1331, #1321, #1320, #1319, #1317, #1344, #1316, #1262, #1139, #1133, #1128, #1131, #1136, #1134, #1253, #1130, #1129, #1127, #1126, #1125, #1256, #1255, #1124, #1123, #1122, #1121, #1120, #1119, #1135, #1257, #1254, #1246, #1092, #1090, #1093, #1091, #1089, #1088, #767, #717, #766, #773, #673, #675, #677, #674, #557, #552, #546, #554, #550, #433, #559, #544, #543, #541, #542, #439, #442, #441, #440, #438, #437, #436, #435, #434
serengil/deepface	23,667	#2 2026-11-10	33	33	33	#1703, #1702, #1696, #1686, #1689, #1685, #1700, #1701, #1684, #1678, #1680, #1679, #1677, #1676, #1675, #1658, #1655, #1651, #1653, #1652, #1649, #1654, #1648, #1646, #1647, #1643, #1644, #1634, #1638, #1636, #1633, #1631, #1630
recommenders-team/recommenders	21,985	#46 2026-11-10	6	6	3	#2419, #2388, #2389
milind-soni/OpenMausBot	20,410	#39 2026-11-10	14	14	8	#5092, #5091, #5090, #5089, #5088, #5086, #5084, #5083
huggingface/sentence-transformers	19,275	#29 2026-11-10	6	6	6	#4323, #4114, #4301, #4127, #4116, #4113
carla-simulator/carla	14,582	#79 2026-11-10	3	3	3	#10191, #10188, #10189
colmap/colmap	13,193	#25 2026-11-10	5	5	5	#4976, #4977, #4975, #4931, #4932
kornia/kornia	12,367	#73 2026-11-10	7	7	7	#6581, #6580, #6582, #6449, #6451, #5517, #5474
lucidrains/denoising-diffusion-pytorch	10,726	#9 2026-11-10	1	1	1	#373
tile-ai/tilelang	8,941	— 2026-11-10 16:39 · retained	—	1	1	#3357
JetBrains/thinkrail	8,070	#16 2026-11-10	3	3	3	#1184, #1186, #1185
openMVG/openMVG	6,596	#96 2026-11-10	1	1	1	#2445
KevinMusgrave/pytorch-metric-learning	6,342	#5 2026-11-10	12	12	8	#779, #778, #777, #776, #775, #756, #759, #758
Eigenwise/atomic-agents	6,321	#23 2026-11-10	2	2	1	#292
ranxianglei/billion-context	6,018	#6 2026-11-10	12	12	10	#6050, #6049, #6052, #6054, #6046, #6053, #6055, #6048, #6047, #6051
rasbt/mlxtend	5,193	#60 2026-11-10	1	1	1	#1207
huamidev/GestaltX	4,079	#2 2026-11-10	13	13	10	#40, #39, #35, #36, #38, #37, #34, #33, #32, #31
mikedh/trimesh	3,719	— 2026-11-10 16:42 · retained	—	0	1	#2618
petercorke/robotics-toolbox-python	3,579	#7 2026-11-10	14	11	11	#801, #800, #799, #798, #797, #796, #795, #794, #793, #792, #719
leggedrobotics/rsl_rl	3,126	#4 2026-11-10	4	3	3	#244, #240, #239
elalish/manifold	2,401	#34 2026-11-10	3	3	3	#1927, #1929, #1905
pypose/pypose	1,621	#16 2026-11-10	6	6	4	#437, #415, #416, #417
geomstats/geomstats	1,529	#83 2026-11-10	1	1	1	#2131
Denys88/rl_games	1,410	#7 2026-11-10	4	4	4	#394, #393, #392, #384
Toni-SM/skrl	1,105	#3 2026-11-10	5	5	5	#489, #496, #469, #467, #475
geoopt/geoopt	1,103	#2 2026-11-10	13	13	13	#280, #279, #278, #254, #253, #255, #256, #258, #259, #257, #267, #266, #262
dfki-ric/pytransform3d	708	#17 2026-11-05 11:26 · retained	2	2	2	#389, #383
naver/roma	647	#2 2026-11-05 11:26 · retained	4	4	4	#33, #32, #30, #29
lucidrains/rectified-flow-pytorch	491	#6 2026-11-05 11:26 · retained	1	1	1	#50
neka-nat/kinpy	181	#3 2026-11-05 11:27 · retained	3	3	3	#29, #28, #27
danijar/elements	73	#3 2026-11-05 11:27 · retained	2	2	2	#18, #19
stephane-caron/qpmpc	56	#2 2026-11-05 11:28 · retained	4	5	3	#17, #16, #15

TD-MPC2’s cherry-picked contribution is recorded under Other accepted work.

Changes since 10 November 2026 · 16:44 CST (UTC+8)

Measure	Previous	Current
Verified merged author PRs	395	397
Merged-PR repositories	36	36
Accepted-contribution repositories	37	37
Default-branch author commits	418	420
Merged PRs in projects above 10,000 stars	300	302
Separate integrations/cherry-picks	15	15
Homepage-eligible upstream projects	30	30
Visible top-30 projects with directly merged PRs	17	17
Personal software stars	17	17
Personal software default-branch commits	157	157

The personal-project comparison uses the previously published 10 November 16:44 project snapshot. Current project record.

Current graph changes

Project	Previous position / graph commits	Current position / graph commits
serengil/deepface	#3 / 31	#2 / 33
huggingface/sentence-transformers	#28 / 6	#29 / 6
carla-simulator/carla	#71 / 3	#79 / 3
colmap/colmap	#28 / 5	#25 / 5
kornia/kornia	#77 / 7	#73 / 7

New directly merged author PRs

Merged (Asia/Shanghai)	Project / PR	Final merge commit	PR title
2026-11-10 18:51	serengil/deepface #1703	15a1e75	Preserve per-face result positions when ANN neighbours fail verification
2026-11-10 18:48	serengil/deepface #1702	30c08b8	Resolve omitted thresholds in dictionary searches

Other accepted work

The following 15 maintainer integrations and cherry-picks have separately recorded adopting commits and attribution. The directly merged author inventory contains 397 PRs. Individual adoption dates remain in the table and source snapshot. Adoption sources.

Project / original PR	Adoption evidence	Attribution
TD-MPC2 #98	Implementation cherry-pick; the authored commit remains in the verified default-branch history. Target-Q history and evaluation mode are preserved during model conversion.	Commit author is Afloat16. This separately adopted contribution brings the accepted-project total to 37; current stars: 999 retained graph position #6 / 1 commit contribution, read 2026-11-05 11:23 Asia/Shanghai.
Robotics Toolbox #735	Adopted by merged #883, 22 October 2026; squash commit. Circular particle-heading moments.	Maintainer-authored squash retains Co-authored-by: Afloat <132809511+Afloat16@users.noreply.github.com>.
Robotics Toolbox #736	Adopted by merged #805, 16 October 2026 (Asia/Shanghai); squash commit. Normalized-time quintic trajectories and the Hermite alternative.	Maintainer-authored squash retains the same Afloat co-author trailer.
Craft Anything #591, #592, #593, #594, #595, #596	Adopted by merged #645, 2026-11-03 20:58 (Asia/Shanghai); commit 0eea78d. HTML scanner correctness for tag boundaries, raw-text elements, preload token lists and URL character references.	Integration PR and commit author: bflatastic; retains the Afloat co-author trailer. The 6 source PRs are Afloat16-authored and were closed without direct merges.
Craft Anything #597, #598, #600, #601	Adopted by merged #649, 2026-11-03 21:07 (Asia/Shanghai); commit 3ee1642. Local preview request scoping, streaming file responses and bounded connection handling.	Integration PR and commit author: bflatastic; retains the Afloat co-author trailer. The 4 source PRs are Afloat16-authored and were closed without direct merges.
Craft Anything #602, #603, #604, #605	Adopted by merged #648, 2026-11-03 21:25 (Asia/Shanghai); commit 82856a3. HAR credential redaction, decoded-byte accounting and valid-date path templates.	Integration PR and commit author: bflatastic; retains the Afloat co-author trailer. The 4 source PRs are Afloat16-authored and were closed without direct merges.
Craft Anything #606, #607, #608, #609, #610, #611, #612, #613, #614, #615	Adopted by merged #652, 2026-11-03 21:29 (Asia/Shanghai); commit 15de366. Binary plist, OID, base64, source-map and archive parser correctness.	Integration PR and commit author: bflatastic; retains the Afloat co-author trailer. The 10 source PRs are Afloat16-authored and were closed without direct merges.
Craft Anything #616, #617, #619, #620, #621	Adopted by merged #653, 2026-11-03 21:41 (Asia/Shanghai); commit aeb86a6. Agent availability and Windows discovery, chat-model gating, bounded SSE events and retry request preservation.	Integration PR and commit author: bflatastic; retains the Afloat co-author trailer. The 5 source PRs are Afloat16-authored and were closed without direct merges.
Craft Anything #640, #641, #642	Adopted by merged #656, 2026-11-03 21:49 (Asia/Shanghai); commit ebed13d. Conditional ARM32 call dataflow, esbuild dependency discovery under caps and distinct window-comparison labels.	Integration PR and commit author: bflatastic; retains the Afloat co-author trailer. The 3 source PRs are Afloat16-authored and were closed without direct merges.
Craft Anything #676	Adopted by merged #683, 2026-11-04 01:33 (Asia/Shanghai); commit 636b3ff. Bound CRC-CPIO checksum work across firmware probes, with shared byte limits, cancellation checks and explicit skipped-verification results.	Integration PR and commit author: bflatastic; retains the Afloat co-author trailer. The source PR is Afloat16-authored and was closed without a direct merge.
Craft Anything #688, #690, #691, #692, #693, #694, #695, #696, #697, #698, #699, #700, #701, #702, #703, #704, #705, #706, #707, #708, #709, #710, #711, #712, #713, #714, #715	Adopted by merged #722, 2026-11-04 02:43 (Asia/Shanghai); commit b9ca099. Respect declared lengths and reject invalid values across bytecode, database, archive, filesystem, configuration and packet parsers.	Integration PR and commit author: bflatastic; retains the Afloat co-author trailer. The 27 source PRs are Afloat16-authored and were closed without direct merges.
Claude-Mem #5339	Extra invalid-input regression cases adopted into merged #5336, 2026-11-04 11:18 (Asia/Shanghai); commit de134d3.	PR and final-commit author: ryrenz; maintainer thedotmack integrated the extra tests and retained the Afloat co-author trailer. Only the rejection tests are credited to Afloat; #5339's production implementation was not adopted.
Craft Anything #842, #844, #877, #878, #879, #880, #881, #882, #883, #884, #885, #886, #888, #889, #890, #891, #892, #893, #894, #895, #896, #897, #898, #899, #900, #901, #902, #903, #904, #905, #906, #907	Adopted by #986, 2026-11-07 00:10 (Asia/Shanghai); commit 60409b6. Bounded binary-format parsing: declared-length and value checks for network captures, archives, Unity/Unreal, binaries and other formats.	Integration PR and final commit: bflatastic. The final squash includes an Afloat trailer. The maintainer combined overlapping DNS fixes from #842, #880 and #881; the ZIP padding contribution from #975 is by Silvertree2010.
Craft Anything #716, #718	Adopted by #995, 2026-11-07 00:50 (Asia/Shanghai); commit e0922f4. Exclusive creation of atomic-write temporary files and preservation of distinct HTTP replay samples.	Integration PR and final commit: bflatastic. The final squash has no Afloat trailer. The author retained the source commits; #1012 later restored attribution separately. Pasted-picture naming in #733 is by HerrnJung.
Craft Anything #1140, #1141, #1142, #1143, #1144, #1145, #1146	Adopted by #1217, 2026-11-09 01:55 (Asia/Shanghai); commit cb0bb9d. Bound source-map, ZIP, ASAR, HAR and .NET bundle parsing; launch an installed Python console script and resolve Godot/Electron project paths.	bflatastic authored the integration PR and final commit, which retains Afloat co-authorship. All seven closed source PRs’ exact head SHAs occur in the integration commit list. The maintainer reconciled overlapping status text and relocated a Windows-sensitive test module.

Craft Anything accounts for 11 integrations covering 101 original PRs. The original integration reviews and individual attribution details remain in the previous record. Historical attribution evidence.