fix(state): land every row of a nested list replacement (#274) - #276

Closes #274.

K1 — replacing a nested list with a longer array

state.$resolve("groups.*.items", [0], [a, b]) over a one-row list landed only 0.0. $watch and $scan never saw 0.1, while the DOM rendered it.

• All three write paths in setByAddress.ts run the dependency walk (notifyWrite) before commitWriteCache.
• groups.*.items contains a wildcard, so it is cacheable. The walk read the pre-write array from the cache, the diff against the baseline said "unchanged", and only the old rows were expanded.

Fix, as proposed in the issue. notifyWrite dirties the written address's own cache entry before the walk, for cacheable writes only. $postUpdate already works in that order. commitWriteCache stores the value again right after.

Side effect. The walk now builds the new array's ledger, so a cold $resolve into that nested list right after a structural write no longer throws ListIndexes not found. That characterization test in integration.recursionKnownDefects.test.ts is flipped to a fixed contract.

K2 — a row written and then removed in the same job

Given [a, b, c], writing b and then assigning items = [a, c] folded c's unchanged value with b's prev.

Not taken: dropping every replaced row, as the issue suggested. An element write into list.* (the swap path) puts a row without a value into the ledger without a diff, and the replaced row is not retired. Dropping every replaced row would lose that position's only landing.

Taken: only rows that a list diff retired (isRetiredListIndex) become gone.

$watch gets the same narrowing

$watch had no narrowing at all. Measured before this change:

• The narrowing moved from scanRuntime.ts to watch/rowLanding.ts, and $scan and $watch share it.
• To keep plain row-value batches cheap, $watch looks positions up only for an entry that has a hit on a retired row, or two hits on the same position. Hits are sorted, so those two are adjacent.
• If reading a position throws, that entry's hits do not fire and are reported once as evaluate. The other watches still run.

Tests

• scan.from.test.ts: both DEFECT(#274) tests are flipped (K1 → ["0.0:5", "0.1:6"] and two $watch calls; K2 → []). Two tests are added for element writes into list.*.
• New watch.rowLanding.test.ts has 9 tests. They cover K2, shortening one row or several, emptying, replacing, a moved row next to a removed row, an element write, an element write next to a row write, and a throwing position lookup.
• integration.recursionKnownDefects.test.ts: case (ii) is flipped.
• The state: replacing row objects while keeping their children arrays leaves that row's aggregate stale (X2) #256 risk tests the issue listed pass unchanged: integration.sharedListRowRevival, divergentLedgerReplacement, recursionKnownDefects and bindComponentNestedFor.

Performance

A/B was measured in one process, alternating runs, on happy-dom:

• The synchronous write path for 2,000 row writes went from 1.876 ms to 1.936 ms.
• Measured end to end with the drain, row writes and a 200-row append (with and without $watch) differed within noise, and not in a consistent direction.
• The browser js-framework-benchmark was not run.

Verification

e2e runs in CI. This PR edits the same CHANGELOG.md spot as #273, so whichever merges second needs a trivial rebase.

🤖 Generated with Claude Code