spec: #992 is not a connect bug — four mechanisms eliminated, two now stale

[artokun]

Investigation of #992. Ships a document, not a fix — deliberately.

Refs #992 — still OPEN, blocked on the reporter since 2026-08-10.

Until now this branch was a single empty commit and the whole investigation lived in this
PR description, so nothing in the repo recorded it and merging would have changed nothing. It
now adds docs/design/992-connect-link-drop-eliminations.md, which labels every claim by
how it was established — measured on a rig, or reasoned from source.

Reported

panel_connect returned success for a link that was gone by queue time
(Required input is missing (latent_image)), on a VAEEncode.LATENT fan-out to three
KSamplers. The first of the three vanished. Reporter on panel 0.11.43,
ComfyUI 0.31.1, frontend 1.48.7.

Measured on a live canvas (the reporter's ComfyUI/frontend build)

Against the panel at the branch point (0.11.90), on 2026-08-10:

#	mechanism	links after
1	the reported sequence: 3 connects, then 4 set_widget on KSampler #1	3/3 intact
2	reapplyDefsToLiveNodes	0 changed, 3/3 intact
3	refreshComboOptionsFromDefs	0 changed, 3/3 intact
4	app.registerNodesFromDefs (what panel_refresh_nodes drives)	3/3 intact

Sizes stayed 270x262 throughout. The reporter's affected node went 270x262 → 392x286 → 392x262. None of the four resizes a node, so none reproduces the only fingerprint.

Re-checked 2026-08-14 — two eliminations went stale

main moved 189 commits touching web/ since the branch point.

• Elimination 2 is STALE. 67e05b53 (fix(1172), landed 2026-08-14) added a
  refreshComboOptionsFromDefs call inside reapplyDefsToLiveNodes; the commit's own
  comment says the sweep previously "never touched options.values". The function that was
  measured is not the one that ships. The conclusion probably survives — combo rebuilding
  writes w.options.values, never node.inputs or graph.links — but that is now
  reasoning that replaced a measurement.
• Elimination 4 is transitively affected — panel_refresh_nodes runs
  registerNodesFromDefs and the changed sweep in one pass.
• Elimination 3 holds — the signature gained a mergedInputs per-type cache (The refresh run budget is thin: refreshComboInNodes alone measures ~4.8s of a 9s run #1193); the
  widget loop is unchanged.
• Elimination 1 should be re-run — set-widget.js changed on 2026-08-11 (fix(1087)).

The reframing — reasoned from source, now pinned to the reporter's build

graph_connect does not report success on LiteGraph's return value; that was #397 and it
was fixed. isLinkPersisted requires graph.links[link.id] to exist and the target input
to back-reference that same id, synchronously, failing closed — and the success payload is
unreachable unless it passed. So the payload was accurate when issued: the link existed and
was destroyed later. The reporter's retry succeeding and persisting fits.

The original version of this claim compared against current source, which does not establish
what the reporter was running. Now measured by blob identity: web/js/lib/connect-verify.js
is blob 5fae91a2 — byte-identical in 0.11.43 (release commit a1ac583f), at the
merge-base, and on main — with the isLinkPersisted call site wired in 0.11.43.

Trap recorded in the doc: git tag --contains reports the earliest containing tag as
v0.11.83 and so reads as "absent from 0.11.43". This repo has 21 tags and tagging began at
v0.11.83, so 0.11.43 has no tag at all. Settle version questions by ancestry against the
release commit.

A fifth mechanism, ruled out from the report itself

LiteGraph destroys the wire a connect replaces; the panel discloses that as replaced_link,
and the disclosure existed in 0.11.43. 142.latent_image was connected exactly once and that
payload carried no replaced_link — so replacement-on-reconnect did not destroy this link.

Why this ships no fix

Hardening panel_connect with a re-verification pass would be ceremony aimed at the wrong
mechanism: it would leave the actual link-destroyer in place, slow every graph build, muddy the
tool's contract, and let the issue be closed as addressed.

Still blocked on the reporter for the panel_edit_node resize calls on 141/144/147 — a resize
is the only operation in their sequence that changes node geometry, which is the one
fingerprint left.

Verification

• Merged origin/main (clean, no conflict markers).
• npm run test:unit: all four gates green (i18n, i18n-render, tool-vocabulary, panel-scope);
  4397/4399 pass, 1 todo. The single failure is manager-install.test.mjs "panel_install_node times out while the canvas remains responsive #671
  verifyInstalled … inside the reply window", a wall-clock budget assertion that measured
  19212 ms under full-suite parallel load; the file passes 122/122 run alone. This branch
  changes no executable code.

[Copilot]

Copilot wasn't able to review any files in this pull request.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.