Spike: remove trackedData and other copies of value-plus-tag state

[NullVoxPopuli-ai-agent]

Spike: one way to hold reactive state. trackedData is gone, and four other copies of "a value next to a tag" now use TrackedValue or the tag registry.

• 14 files changed, 125 lines added, 327 removed, against the content of Store each @tracked field in a TrackedValue #21662.
• After the rebase, the full suite passes with legacy decorators: 9522 pass, 18 skipped, 0 failed. Before the rebase it also passed with VITE_STABLE_DECORATORS=true. That mode did not run again.
• After the rebase, against main: pnpm bench shows no change (-0.1%, 0 of 23 phases), and rere-benchmark shows -0.2% across 8 cycles, which is no clear change. Against Store each @tracked field in a TrackedValue #21662 alone, the four spike commits are 1.8% slower in rere-benchmark, inside the spread of the cycles.

Depends on #21662

The branch is rebased on main of 2026-10-07. It holds the one commit of #21662 (9f82b544f7), then the four commits of this spike. The parts of #21650 that it needed are merged. The PRs are not a GitHub stack, because the branches are in a fork and the repository does not have stacked PRs on.

The four commits

commit	what was there	now
c93f7db5b4 Remove trackedData	tracked-data.ts: a WeakMap of values plus the tag registry. It had no caller after #21662.	deleted, with its export and its tests
f00920fb3b Keep a tracked accessor in a TrackedValue too	the accessor form of @tracked had its own copy of the tag logic, with a registry lookup on each read and write	the storage of the accessor holds a TrackedValue. The field form and the accessor form share one read function and one write function.
5839bbc474 Keep the list item value and the root outlet state in a TrackedValue	createIteratorItemRef and createRootOutletState were each a closure with a value and a tag	a TrackedValue
f45c1cbde8 Use the tag registry functions in the tracked collections	each of the six collections had its own "get or make the tag of a key" and "dirty the tag of a key"	tagFor and dirtyTagFor, with the map of the collection as the store

What changes for callers

• trackedData is no longer an export of @glimmer/validator. In the repo it had one caller. As one sample of outside use, the 1,843 packages in the node_modules of one Ember app have no other user.
• tagFor and dirtyTagFor take any key type, and a store with get and set. A WeakMap fits.
• In debug builds, a write to a tracked collection after a read in the same computation now uses the assertion of dirtyTagFor. The message names the object and the key.
• A tracked accessor makes its TrackedValue, and asks the registry for its tag, when the instance is made. Before, the registry made the tag at the first read.

Looked at and left alone

These also keep a value or a revision next to a tag, but they are not a second way to hold source state:

• createCache, ReferenceImpl, computed properties, and observers keep a derived value with a tag and a revision. ReferenceImpl repeats what createCache does. That is a bigger change than this spike.
• Component[DIRTY_TAG] and Helper[RECOMPUTE_TAG] are tags with no value.
• Modifier tags and the lazy chain tags are tags that updateTag points at other tags.
• CUSTOM_TAG_FOR computes a tag on demand for args and proxies.
• smoke-tests/benchmark-app/app/services/cell.js is a copy of TrackedValue. I left it, because a change there changes the control side of pnpm bench.

Measurements

After the rebase

Control is main at 9bec1cb2a8. The spike is f45c1cbde8, and #21662 is 9f82b544f7.

pnpm bench, 50 paired rounds, spike against main: tracerbench-report.pdf

• Script time for the full run without GC: -0.1%, which is not significant. 0 of 23 phases have a significant change.

rere-benchmark, 8 mirrored cycles, 16 runs for each build, 0 void runs, geometric mean of 13 benches:

pair	change	c1	c2	c3	c4	c5	c6	c7	c8
#21662 against main	-1.9%	+0.7%	-4.2%	+3.4%	+2.3%	-7.5%	-2.7%	-0.3%	-0.7%
this spike against main	-0.2%	-0.3%	-2.9%	+1.8%	+8.7%	-5.2%	+0.4%	-3.4%	+5.1%
this spike against #21662	+1.8%	-1.0%	+1.4%	-1.5%	+6.3%	+2.4%	+3.2%	-3.0%	+5.8%

• The main build drifted 5.4% between the first four cycles and the last four, and the sign of each pair changes between cycles. No pair shows a clear change.
• The spike is slower than Store each @tracked field in a TrackedValue #21662 in 5 of 8 cycles. The largest single bench is 1k items 1 update on 5% (random, async) at 11.9% slower. A cost of the TrackedValue for each list item is possible, but this round does not prove it.
• 1 item, 100k updates: Store each @tracked field in a TrackedValue #21662 is 11.8% faster than main, and the spike is 9.5% faster.

Before the rebase

These numbers are from the head before the rebase (249ec6cef0). The control for the four spike commits is the content of #21662 on top of #21650.

pnpm bench, 50 paired rounds, script time without GC: tracerbench-report.pdf

• Full run: -0.6%, which is not significant.
• 3 of 23 phases have a significant change: render1000Items2 is 3.9% faster, render10000Items2 is 1.3% faster, and swapRows1 is 5.0% slower.

rere-benchmark, 4 mirrored cycles, 8 runs for each build, 0 void runs, geometric mean of 13 benches:

pair	change	cycle 1	cycle 2	cycle 3	cycle 4
#21662 against main	-4.6%	-6.3%	-3.8%	-1.2%	-6.6%
this spike against main	-3.1%	-7.2%	-2.8%	+1.3%	-2.6%
this spike against #21662	+1.7%	-1.0%	+1.0%	+2.5%	+4.2%

• The four spike commits cost between nothing and a few percent. The sign changes between cycles, so this round does not prove a cost.
• One bench stands out: 1k items 1 update on 5% (random) is 17.2% slower than Store each @tracked field in a TrackedValue #21662. Its async twin is 3.1% slower, and the 25% version is 5.6% faster, so this can be noise.
• The small cell of Keep the value and the tag of a tracked field in one cell #21661 stays the fastest option for a tracked field: Store each @tracked field in a TrackedValue #21662 is 3.9% slower than the cell in an earlier round.

Tests

• The trackedData tests are deleted. The decorator tests cover the same cases: a first read with and without an initializer, a write before a read, and the tag of the field.
• tsc --noEmit, ESLint, Prettier and pnpm lint:docs pass.

🤖 Generated with Claude Code

[NullVoxPopuli]

this is kinda messy right now. other PRs need to land for this to look like a smaller PR

[NullVoxPopuli]

this package is not public, so 🙈