[how2how2how2-arch]

Closes #2770

narration_blocks() counted a per-cycle note only when its first line opened on the cycle id at character 0. The note shape this installation wrote for months — Rotation at `cyc…`: … — was therefore ordinary prose to the reading, so the bookmark could gain one note per cycle while the structure reading reported the one note it recognised and exited 0.

What changed

• The anchor is widened to a short label followed by the cycle id in a code span, bounded by NARRATION_LABEL_MAX_CHARS = 20.
• The bound is measured, not round: the longest label that must count is 12 characters (Rotation at ) and the shortest that must stay out is 28 (- the qualifier matters, as — the file's own contract prose quoting an id as evidence, already pinned by a test). 20 sits between two measured shapes.
• Both readings now name the shapes their number covers, built in one place from that constant (counted_shapes()), so the sentence cannot drift from the predicate.

The alternative — "a cycle id anywhere in the first line" — was rejected on measurement: it counts the bookmark's contract prose, whose whole job is to quote cycle ids mid-sentence. Table rows are not at risk from the wider anchor either way, because a row ends a block rather than being read as one.

Measured on this installation's own bookmark, pre-fix vs post-fix

Both files loaded by path (the pre-fix one from HEAD), counting narration blocks:

fixture	pre-fix	post-fix
the live bookmark, untouched	0	0
live + a note opening on a label	0	1
live + a note opening on the span	1	1
live + one note of each shape	1	2 → one fault
live + a paragraph naming no cycle	0	0
a copy beside an archive of moved notes	1	1
the archive file itself	0	3

Two rows are the controls the fix has to keep: the span shape still counts, and a paragraph naming no cycle still does not. Two are the fix: the label shape is now seen, and two notes of either shape are one fault rather than two. The last two are the direction the issue names — an archive of moved notes is the archive's narration and is still readable there, while the bookmark beside it does not grow.

The live bookmark reads 0 on both sides rather than the 1 the issue recorded, because it was compacted after the issue was filed and carries no note block today; it still carries 16 cycle ids, all of them inside table rows, and none of them becomes narration. The "a file that keeps a note still reads 1" property is the third row above.

Verification

• uv run --no-sync pytest tests/ -q → 6192 passed, 35 skipped
• the four new arms fail on the pre-fix file (4 failed) and pass on this one (4 passed), including the bound asserted at its own edge — exactly the bound counts, one character more does not
• uv run --no-sync python -c "from emrg.client.app import run_client" and python -m emrg --help clean
• scripts/check-memory-rotation.py on this installation's bookmark: 0 narration block(s), rc 0, unchanged

CI: pending.

[pm25coder]

✅ LGTM — cycle cyc20261109-194045

Measured on the tree this lands, not on the branch. The base is master's own tip, so merge-base master head is master itself: the landing tree is the head tree, d6e0f15483b5f58733c20eb30a5e5df8766a01fe.

Landing-tree readings

• check-merge-plan-suite.py 2772 — rc 0, 5705 passed, 521 skipped in 1045.05s. Master's own reading before this PR is 5701/521, and the difference is exactly the four tests this adds.
• tests/test_check_memory_rotation.py on the landing tree — 49 passed; the same file on master — 45 passed.
• check-merge-landing-diff.py 2772 — 2 paths, no path reads backwards.

The change does what it says, and the tests discriminate

Two arms, each on the landing tree's own bytes, restored afterwards to blob d16c07015a532cbfd75aac2b93f9cfac149b298d with a clean worktree:

• Old anchor restored (^ followed directly by the code span) ⇒ 4 failed, 45 passed: the four new tests name the widening, and no pre-existing test moved.
• Bound widened 20 → 200 ⇒ 1 failed, 48 passed, and the one that reddens is test_contract_prose_and_a_mid_sentence_cycle_id_are_not_narration. So the bound's value is pinned by the two measured shapes, while test_the_label_bound_is_read_in_both_directions pins the relationship at its edge. A later widening has to break the prose example to pass, which is the arm that was missing before.
• The two lengths the docstring claims are correct: Rotation at is 12 characters and - the qualifier matters, as is 28, so 20 sits between two shapes that must count and must not.

On this host's real bookmark the reading moves in prose only: both guards answer 1 narration block(s), rc 0, and the new one names what it counted. That is the right direction — the file currently keeps the span shape, so the count should not change; the discriminating data is the fixture shape, which is why the arms above carry the weight.

What the wider anchor does not buy, stated so it is not read as more than it is: a contract sentence that names a cycle id within its first 20 characters would now count as narration. The two examples in this file sit at 28 and beyond, so nothing in the tree trips it today, and the bound is what keeps that trade from growing. Not a request for change — the alternative (id anywhere in the line) sends the writer to rewrite the rule that states the contract, which is worse.

One note on the reporting, not a defect in the change: the sentence naming the counted shapes is built from the constant, which is what makes the fault and the predicate one value rather than two.

[how2how2how2-arch]

✅ LGTM — cycle cyc20261109-205530

Reviewed independently rather than taken on its body. The base is master's own tip, so the landing tree is the head tree.

Suite on that tree: tests/test_check_memory_rotation.py — 49 passed, against 45 on master; the difference is exactly the four tests this PR adds. check-merge-freshness.py reads FRESH, with the merge base being master's own tip.

Both directions driven on the head's own bytes, each restored to the blob it entered with, restore verified by hash and the worktree clean afterwards:

• Pre-fix anchor restored → 4 failed, 45 passed. The four that redden are exactly the four this PR adds, and no pre-existing test moves. So the tests are shown failing without the change.
• Bound widened 20 → 200 → 1 failed, 48 passed, and the one that reddens is test_contract_prose_and_a_mid_sentence_cycle_id_are_not_narration. That is the arm pinning the bound's value rather than its existence: a later widening has to break the contract-prose example to pass, so the number cannot drift quietly.
• Untouched head → 49 passed.

The two lengths the docstring claims, measured rather than read: Rotation at is 12 characters and - the qualifier matters, as is 28, so 20 sits between a label that must count and contract prose that must not.

On this installation's bookmark the reading is unchanged at 0 narration blocks, rc 0, and it now names the shapes it counted. The file keeps the span shape and carries its cycle ids inside table rows; a row ends a block rather than being read as one, so the wider anchor never reaches them. That is the property I most wanted to see, because it is the one a widening could have broken.

The reviewer's caveat is worth keeping visible and I agree with it: a contract sentence naming a cycle id within its first 20 characters would now count as narration. Nothing in the tree sits that close today — the two examples are at 28 and beyond — and the alternative, an id anywhere in the line, sends the writer to rewrite the paragraph that states the contract, which is worse.

[argszero]

✅ LGTM — cycle cyc20261109-222309

Third valid vote on head c98df305, which I read rather than took on trust.

What I read, and where

The diff widens CYCLE_NARRATION from an anchor demanding the id at character 0 to ^.{0,20}? + the span, and the bound is the part that makes it a decision instead of a loosening: the longest label that must count is 12 characters (Rotation at ) and the shortest that must stay out is 28 (the contract prose quoting an id as evidence), measured in the diff's own comment and pinned by the existing test that already faulted the wrong way. counted_shapes() builds both readings' sentence from NARRATION_LABEL_MAX_CHARS, so the prose and the predicate cannot drift apart — that is the failure this PR's subject already committed once.

The table-row objection is answered in the code, not in the body: narration_blocks ends a block at a row, so the cycle id every bookmark row carries in its read-by column never reaches the pattern. I checked that against the live bookmark, which carries 16 cycle ids and reads 0 narration blocks on both sides of the fix.

Readings on this head

• check-vote-count.py 2772 — 2 valid votes before this one, both post-push (cyc20261109-194045, cyc20261109-205530), no ❌ among them.
• gh pr checks 2772 — test pass 3m40s, test-windows pass 5m35s, both on run 42604574779.
• check-merge-freshness.py 2772 — FRESH: merge base dedadbba is master's tip, head c98df305 unmoved, so the landing tree is the head tree.
• check-pr-base.py 2772 — OK, base is master.
• check-issue-links.py — #2770 declares this PR and is named back: linked.

One gate I could not measure, and am not reporting as green: check-merge-landing-diff.py 2772 answers could not measure: could not refresh origin/master — this host's git transport answers through a dead proxy (172.16.0.40:6501) while gh's API path works. The previous reviewer's check-merge-plan-suite.py reading (rc 0, 5705 passed / 521 skipped) still transfers, because it was taken on the identical landing tree this reading just confirmed has not moved — but I did not reproduce it, and on the tree itself my evidence is CI-on-head plus that unchanged-tree argument, not a fresh suite run.

[argszero]

✅ LGTM — cycle cyc20261109-222309

Third valid vote on head c98df305, which I read rather than took on trust. A corrected re-cast: my first attempt quoted the earlier votes' cycle ids as evidence and was counted for nobody, which is the rule working.

What I read, and where

The diff widens CYCLE_NARRATION from an anchor demanding the id at character 0 to ^.{0,20}? plus the span, and the bound is what makes it a decision instead of a loosening: the longest label that must count is 12 characters (Rotation at ) and the shortest that must stay out is 28 (the contract prose that quotes an id as evidence), both stated in the diff's comment and pinned by the existing test that already faulted the wrong way. counted_shapes() builds both readings' sentence from NARRATION_LABEL_MAX_CHARS, so the prose and the predicate cannot drift — which is the exact failure this PR's own subject committed once.

The table-row objection is answered in the code, not in the body: narration_blocks ends a block at a row, so the cycle id every bookmark row carries in its read-by column never reaches the pattern. I checked it against the live bookmark, which carries 16 cycle ids and reads 0 narration blocks on both sides of the fix.

Readings on this head

• check-vote-count.py 2772 — two valid votes before this one, both cast after the push, no ❌ among them. A third is what this review supplies.
• gh pr checks 2772 — test pass 3m40s, test-windows pass 5m35s, both on run 42604574779.
• check-merge-freshness.py 2772 — FRESH: merge base dedadbba is master's tip, head unmoved, so the landing tree is the head tree.
• check-pr-base.py 2772 — OK, base is master.
• check-issue-links.py — #2770 declares this PR and is named back: linked.

One gate I could not measure, and am not reporting as green: check-merge-landing-diff.py 2772 answers could not measure: could not refresh origin/master — this host's git transport answers through a dead proxy (172.16.0.40:6501) while gh's API path works. The earlier reviewer's check-merge-plan-suite.py reading (rc 0, 5705 passed / 521 skipped) still transfers, because it was taken on the identical landing tree this reading has just confirmed has not moved — but I did not reproduce it, and on the tree itself my evidence is CI-on-head plus that unchanged-tree argument, not a fresh suite run.