One-hop tracked storage and known-tag references

[NullVoxPopuli-ai-agent]

Extracted from the #21520 spike; independent of the other extraction PRs. Full suite green locally: 9443 tests, 0 failures — including the notifyPropertyChange and computed-property interop this touches most directly.

One-hop tracked storage: trackedData keeps value and tag in a single cell per (field, instance) — a tracked read is one WeakMap hop + consumeTag, a write is one hop + DIRTY_TAG, replacing three map hops each way (central TRACKED_TAGS WeakMap → per-object Map, plus the separate values WeakMap). The cell registers its tag into the central registry at creation (registerTagFor), so notifyPropertyChange, computed chains, and anything else resolving tags via tagFor/dirtyTagFor observes the identical tag object the field consumes and dirties — interop is preserved by tag identity, not by parallel bookkeeping.

Known-tag references: a child reference locks onto its property's canonical tag when its first framed compute consumed exactly that tag. A single consumed tag rules out branching getters (their consumed sets vary run to run); identity with the registry's cell tag proves the read was the plain tracked-field getter on a parent that cannot change (a mutable parent's tag would have been in the frame too — constants are dropped from frames). Locked references recompute with no frame push, no tag collection, and no combine(); the getter's own consumeTag lands in the ambient frame, which is exactly what the framed path's trailing consume achieved. Anything unproven stays on the framed path forever.

Honest impact framing: neutral on the spike's benchmark suite (its apps are worker-POJO- and helper-bound); the win scales with tracked-getter density and template property reads — the common shape of real apps — and it removes per-read allocation and map traffic unconditionally.

Related: #21520 (spike), #21543 (pools), #21544/#21545 (update-path stack).

🤖 Generated with Claude Code

[NullVoxPopuli-ai-agent]

Isolated same-batch measurement (today's main vs this branch, 8x throttle, mutation-anchored) — regressive on walk-heavy paths; closing under the same standard that closed #21543.

bench	main	this PR
DB Monitor	10.6 fps (10–13)	7.1 fps (7–12) — two of three runs below main's entire range
1k items, 25% random (async)	537ms (401–663)	878ms (598–1230) — four of five runs at/above main's max
1 item, 1k sync / 100k sync	13.1ms / 42.7ms	10.1ms / 36.9ms (the promised set-path win, real but small)
everything else	—	neutral

The sync one-shot improvement is real but tiny in absolute terms, and it's paid for with a genuine regression where ember is weakest. Likely mechanism: ReferenceImpl gained three fields (bigger objects for every ref in ref-dense trees) and every childRefFor does candidacy bookkeeping, while the locking that was supposed to pay for it rarely fires on real templates (list-item parents are mutable by design, so their refs correctly never lock). The cost is unconditional; the win is conditional — backwards.

On the spike branch these changes measured differently (its scheduler coalesces to far fewer flushes, hiding per-flush ref costs), which is exactly why the isolated-on-main standard exists. The one-hop tracked-cell storage might survive on its own without the reference-locking half — if it's wanted, it should return as a separate PR with its own isolated numbers.

🤖 Generated with Claude Code