[SPIKE / not for merge] Push-invalidation: tags queue their consumers, no revalidation walk

[NullVoxPopuli-ai-agent]

Documentation PR — explored and declined. This is the "what if there is no revalidation?" experiment referenced from #21513 (rounds 4-5), given its own diff so the design and its costs are reviewable. It is stacked on #21513 but replaces that branch's revalidation design rather than adding to it — the two are mutually exclusive.

The design (final form)

• tags gain subscribers; dirtyTag notifies them synchronously
• {{#each}} items subscribe to their collected subtree-tag leaves; list blocks collect sync-consumed tags in a tracking frame (iteration reads collection cells lazily — subscribing to the iterable ref alone misses trackedArray.shift)
• render roots render inside a tracking frame and subscribe to what they read; items/blocks do NOT propagate deps upward, so a root re-render fires only when the root's own non-delegated deps change
• no full walk exists: with complete coverage, unsubscribed dirt provably affects nothing rendered and is ignored; item bootstrap flows through the queue (blocks enqueue never-collected children)
• drains coalesce to one requestAnimationFrame

Why declined

Same-batch dbmon at 8x CPU throttle (svelte anchors stable):

configuration	fps	vs svelte
stock 7.3.0-alpha.5	7.2	22%
#21512 subtree-skip	13.7	42%
#21513 coalesced-walk stack	27.1	82%
this branch (walk-free push)	22.3	68%

The theoretical O(changed) win is eaten by per-flush subscription churn (unsubscribe/resubscribe per processed opcode), root bookkeeping, and high run-to-run variance. Two design findings worth keeping regardless:

1. Partial subscription coverage is silently incoherent — un-subscribed regions go stale and can be masked by unrelated fallback renders. Push is all-or-nothing.
2. rAF deferral breaks render-latency-coupled patterns — a set→effect→advance loop went 0.4s → 33.6s (every iteration waits a frame). This caveat applies equally to [SPIKE] Stacked update-path optimizations: how fast can list-heavy updates get? #21513's coalescing lever and needs a sync-flush escape hatch in any landable frame-aligned design.

Conclusion: the walk was never the enemy — per-message flushing and legacy read tracking were. See #21513 for the direction worth landing.

🤖 Generated with Claude Code

[NullVoxPopuli]

This is the branch that got good results

[NullVoxPopuli-ai-agent]

pnpm bench for this branch (walk-free push-invalidation + rAF drains). Tracerbench krausest vs origin/main, 20 samples, per-phase medians + Mann-Whitney p from raw compare.json:

21 of 23 phases neutral; two phases reach significance:

phase	control	experiment	delta
renderEnd (initial render)	28.5ms	30.3ms	+1.9ms (+6.6%) regression (p=0.037)
render1000Items1	139.8ms	119.5ms	-20.3ms (-14.5%) improvement (p=0.006)

Both are coherent with the design rather than noise-shaped: initial render now pays subscription setup (root tracking frame + leaf subscription per item bootstrap), and the first big list render benefits from items skipping the collection walk that the base branch's finalizer does eagerly. The rest — updates, selects, swaps, removes — is flat; krausest's single-interaction shapes don't exercise the sustained-stream scenario where the push-vs-walk difference lives (and where this branch measures worse than #21513: 68% vs 82% of svelte on dbmon).

This completes the bench triptych: #21512 fully neutral, #21513 fully neutral, this branch neutral-with-two-small-signals. None of the krausest results changes the conclusion in the PR description: explored, measured, declined in favor of #21513's direction.

[NullVoxPopuli-ai-agent]

Closing: the push-invalidation question was answered — at best parity with coalesced-walk+subtree-skip at far higher complexity, and the walk was never the bottleneck. Findings are recorded in the round-4/5 analysis comments; the winning direction lives in #21520.