Two separate Relates strongly holding each other's key as their value
(inst->slot in one, slot->inst in the other) is a mutual-reachability
cycle that ordinary weak-key GC can't resolve on its own -- the case
Lua 5.2 ephemerons exist for. Weaken both to pure lookups and make
bindLifetime/unbindLifetime (already used for Slot's own observers)
the single GC anchor for the slot object itself, which attachSlot/
destroySlotTree were missing.
Audited every other Relate() in base/ for the same shape (an object
other than inst used as the outer key) -- slotOwner was the only one;
everything else either has no back-reference to inst or is the safe
single-table self-reference shape bindLifetime already relies on.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Position-keyed comparison only catches "did this exact slot spot change,"
not "is this same Slot object already mounted somewhere else" -- which is
exactly the double-mount invariant Slot already promises for its elements.
A Relate<slot, inst> enforces it directly: same owner -> spurious re-emit,
ignore; different owner -> error; no owner -> bind.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Corrects a foundational error: "handler type unchanged -> retract skipped,
process diffs" was never actually true. StoreBind unconditionally calls
Dispatch.retractUnder before every re-dispatch (already stated in the
original dispatch model doc) -- Tag's old assert(v==nil) was taken at
face value and the general rule was wrongly back-derived from it.
Tag, Ref, Slot, and Attribute (the last three designed earlier this same
session) all inherited the flawed premise. Fixes across all four:
- Tag: rewritten around kTagMap/tagNameMap reference counting so
AddTag is entirely process's job, RemoveTag entirely retract's,
using a Contains hint to skip engine calls that would be immediately
redone. Also fixes the case where two independent Tag(...) values at
different positions share a name (union semantics, like className).
- Ref/Slot: drop the incorrect assert(v==nil), split into "retract
unbinds/destroys the old, process binds/mounts the new" with an
identity check on both sides to avoid spurious re-fire.
- Attribute: AttributeKeyHandler.retract gated on v==nil so ordinary
value updates don't flicker the attribute to nil; the group's
"unchanged name" delegation now also retracts before reprocessing,
since skipping it was silently stacking chain entries every cycle.
Reversal preserved in archive/retract-always-fires-reversed.md.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Group Attribute(...) delegating through the public memoized AttributeKey(name)
cache meant a direct [AttributeKey "name"]=v write and a group field (or two
independent groups) targeting the same name silently converged on the same
dispatch slot -- last write wins, no error. Fix: groups use rawNew(name) to
mint a private per-name key cached on their own (inst,index) relation instead
of the shared cache, so AttributeKeyHandler can detect a conflicting claimant
by simple object-identity comparison -- no separate ownership registry needed.
Persisting names reuse the same cached key across group value swaps, so the
existing diff (nil only what left, leave the rest alone) keeps working.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Same gap as Ref: slotA -> slotB via State<Slot> matches the same
SlotHandler both times, so Dispatch skips retract and process must
diff. Slot's confirmed no-portal/discard-only policy makes an identity
short-circuit a correctness requirement, not just an optimization --
without it, any spurious same-value re-emit would tear down and
rebuild the entire mounted subtree.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
State<Ref> rebinds (refA -> refB) go through process's own diff (same
handler type matches both), not retract -- exactly the TagHandler
precedent. retract only fires when the slot stops being a Ref at all.
Both paths converge on old:Set(nil). Non-nilable Ref<T> remains a valid
"read a settled value" use case; callers opt into Ref<T?> explicitly for
Store/Modifier slots. Unbinding is independent of Instance Destroy --
that's Effect's job, not Ref's.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
PreRef never enters the normal retract dispatch chain (consumed via None
in the pre-pass), so "cancel" was never structurally possible. The real
hazard is reuse across constructions (stale .Value silently firing
callbacks) — guard it with an explicit error on re-fire instead.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Completed 시 per-instance 북키핑 정리는 불필요 — 자연완료는 유저가 원한
목표값에 도달한 상태라 남은 참조가 부작용 없고, Value가 lerp 가능한
프리미티브라 메모리 문제도 없어 별도 정리 장치는 오버엔지니어링. 이걸로
tween-plan.md에 남은 열린 질문이 없어져 research/에서 base/로 승격,
라이브 크로스레퍼런스 전부 갱신.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Sketch Sum/Not/Product-style :Compute/:Apply sugar (research/operator-sugar-plan.md,
implementation deferred). Reusable curried combinators (Sum, Animate) must
go through :Apply, not :Compute — quad rejected implicit auto-tracking, so
a factory's closed-over deps only register if the factory re-declares them
via self:Compute(...) internally; plugging a pre-built factory straight
into :Compute silently drops reactivity. Flip Animate's call site
accordingly in tween-plan.md, add the convention to bind-system-plan.md's
:Apply section, and fix a stale two-arg Animate signature comment in
architecture.md.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
state:Compute(fn)'s fn receives lazy State handles, not raw values, per
the confirmed self/with contract. Found and fixed the same class of bug
the user caught in the Animate example in three more spots (slot-plan.md
x2, tag-plan.md). Added a warning note near the contract since this is
an easy mistake to repeat.
CanAnimate: State<boolean>|boolean|nil (nil defaults true) lets Animate
bypass wrapping in Tween{} entirely, covering the reduceMotion use case
natively. Also documents that Luau's if-then-else expression and const
bindings are official syntax (not hallucinated), so agents don't revert
them; const adoption deferred pending tooling support.
Animate(info) resolves T|State<T> option fields into a plain Tween{...},
matching :Compute's self-as-lazy-handle contract directly. Also audited
.claude/base for and/or ternary idioms per Luau if-then-else guidance;
fixed a real falsy-value leak in Dispatch.retractUnder.
Adds Attribute(store1, store2, ...) as a Tag-shaped array-part value
object that projects one or more Stores' named Source fields onto
native Roblox attributes in one binding, with .Merged for combining
heterogeneous Stores. Rejected: a bare [Attribute] = Store{...} hash
slot (only one slot per instance, can't merge heterogeneous stores)
and making Attribute a Store subtype (would let Store<T>'s T be a
Store again, colliding with the existing "handler-layer values can't
live in Source" rule). Verified the value-slot aggregation in .Merged
doesn't reopen the earlier "layered Store" rejection (archive/
context-rejected.md) since it's an explicit one-time author-time
composition, not an implicit read-time parent-chain fallback.
Renamed the existing single-key constructor Attribute<<T>>(name) to
AttributeKey<<T>>(name) to remove the name collision with the new
group primitive (provisional per existing OnChange/OnChangeKey
precedent; final bikeshed still queued in question.md).
Follow-up: AttributeKey(name) now memoizes through a name-keyed weak
table, guaranteeing AttributeKey(a) == AttributeKey(a) while any
strong reference (e.g. a live Dispatch chain entry) keeps it alive.
This let the group Attribute handler drop its originally-planned
self-contained SetAttribute/subscription logic and instead recurse
into the existing single-key AttributeKey dispatch path per field,
reusing its None/retract/store-bind handling instead of duplicating
it — the group handler's own state shrinks to just a prior-name-set
diff for deciding which keys to retractUnder. The same memoization
was applied to OnChangeKey for the same reason (pure name-to-key
mapping, no other varying state), confirmed safe even once
State<function> callbacks are involved.
Reflected across base/attribute-plan.md, base/onchange-plan.md,
base/architecture.md (source tree + Brand isX list), ROADMAP.md M2/M10,
question.md, README.md, and the luau-test type-narrowing spike.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
CLAUDE.md had grown to 3196 lines of accumulated session logs, causing
context bloat. Full session narratives (including trial-and-error and
later-corrected reasoning — quadnomicon devlog raw material) now live
as 39 individual files under .claude/session/. CLAUDE.md keeps only a
short "지금 할 일" (re-synced against question.md/pre-implementation-audit.md,
stale detail dropped) and a compact per-session summary+link table.
No design decisions changed; base/research/question.md were already
in sync with every session (verified against README.md/question.md
before archiving), so no unreflected content needed migrating first.
