docs(workflow-builder): verify the last 27 register entries — every fix proposal now measured

Completes the verification pass. All 43 C-cluster `fix` proposals have now been
checked against current source; 5 (12%) were already fixed.

Two newly found stale, both fixed by earlier work that cited its own gap id in
a comment — G-038 by G-010's `resolveAnchorTarget`, G-064 by G-012's shared
`noOutputReasonForNode`.

Measured staleness is 12%. The 33% I estimated from a four-entry sample was
wrong in the safe direction but was still an extrapolation and should not have
been offered as a rate — C1 ran 18%, C3 ran 0%.

The pass also surfaced a category the register has no column for: entries that
are structurally true but that nothing shipped can reach.

- G-066 — `registerArtifactKind` has ZERO production call sites, so the stored
  value the frozen snapshot would destroy cannot exist.
- G-046 half one — ZERO of the catalog's activities declare a kindless input
  port. Five unreachable binding states, for an empty port family.

And the inverse, in the same entry: G-046's other half covers 26 optional
base-`Artifact` ports, each owning a canvas handle invisible to the Inputs
panel, badge and drawer — the same population behind the agent scenario-1
catalog-vs-runtime mismatch. One entry, one empty half and one 26-port half;
they need separate rulings.

Finding those three needed enumeration of what the catalog and call graph
actually contain. Grepping the cited evidence line alone would have confirmed
all three as "still true" and been useless.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: alex-struk

feature-docs/20260724-workflow-builder-spec-completion/DECISION_SHEET.md

@@ -239,7 +239,70 @@ warning (`8fb19d57`, earlier this session) had broken a backend
 that shipped; the backend one was not. Recorded here rather than buried in the
 commit, because the lesson is about which suites a package-level change reaches.
 
-## Remaining verification
+## Verification complete — all 43 C-cluster entries measured, 2026-07-27
+
+The remaining 27 are verified. Combined with C1 and the earlier C5 pass, **every
+`fix` proposal in the register has now been checked against current source.**
+
+| Cluster | Verified | Already fixed | Still true |
+|---|---|---|---|
+| C1 reference integrity | 11 | 2 (G-030, G-048) | 9 → **all 7 remaining now shipped** |
+| C5 control-flow & registry | 11 | 1 (G-067, this session) | 10 |
+| C2 validation surfacing | 7 | **1 (G-038)** | 6 |
+| C4 run observability | 8 | **1 (G-064)** | 7 |
+| C3 composition & authoring | 6 | 0 | 6 |
+| **Total** | **43** | **5 (12%)** | **38** |
+
+**Measured staleness is 12%, not the 50% the first four-entry sample suggested
+and not the 33% I estimated from it.** The estimate was wrong in the safe
+direction, but it was still an extrapolation from four entries and should not
+have been offered as a rate. The clusters differ sharply — C1 ran 18%, C3 ran
+0% — and the only reason to know that is to have measured each.
+
+### The two newly-found stale entries
+
+- **G-038** (workflow-level validation rows inert) — fixed by G-010.
+  `resolveAnchorTarget` now routes edge, group, entryNodeId, ctx and
+  library-port anchors; only the genuinely workflow-level ones stay inert. The
+  fix commit says so in a comment, which is why this was cheap to confirm.
+- **G-064** (wire-peek blames the cache for a producer that never ran) — fixed
+  by G-012. `WirePeekPopover` reads `producerStatus` from `nodeStatuses` and
+  shares `noOutputReasonForNode` with the node card, so the two surfaces agree.
+
+### Three entries are TRUE but not LIVE — this changes what they are worth
+
+Verification turned up a distinct category the register does not have a column
+for: the code is exactly as described, but nothing shipped can reach it. These
+should not be ruled `fix` on the same footing as the rest.
+
+| Entry | Structurally true | But |
+|---|---|---|
+| **G-066** `KindSelect` reads the frozen registry snapshot | yes — `Object.keys(ARTIFACT_REGISTRY)` | `registerArtifactKind` has **zero production call sites** (tests only). No kind can be dynamically registered today, so the value it would destroy cannot exist. |
+| **G-046** half 1 — kindless input ports | yes — `resolveInputPort` returns `unsatisfied` on `kind === undefined` | **0 of the catalog's activities declare a kindless input port.** Five unreachable binding states, for an empty port family. |
+| **G-081** `ctx-bound` unmodelled | yes | already downgraded to minor in pass C for the same reason: `computeNodeStatus` has no production caller. |
+
+**G-046's other half is the opposite — much bigger than the entry implies.**
+There are **26 optional base-`Artifact` input ports** across the catalog
+(`file.prepare.fileName` / `.fileType` / `.contentType`, `azureOcr.poll.modelId`,
+`azureOcr.extract.fileName`, …). Every one owns a canvas handle you can drag
+onto while being invisible to the Inputs panel, the badge and the drawer. These
+are the same ports behind the agent scenario-1 catalog-vs-runtime mismatch, so
+this is a live gap with a known second symptom, not a theoretical one.
+
+The lesson is the same one the linter's shape-coverage pass taught: *a check
+nothing can exercise and a gap nothing can hit are the same kind of finding.*
+Both halves of G-046 were written as one entry because they share a code path;
+they should be ruled separately because one is empty and one is 26 ports wide.
+
+### Method note
+
+Every check was a direct read of the cited evidence line against current source,
+plus — where the claim was about reachability rather than code — an enumeration
+of what the catalog or the call graph actually contains. The second kind found
+all three latent entries; grepping the cited line alone would have confirmed
+every one of them as "still true" and been useless.
+
+## Remaining verification (historical — now complete)
 
 C2 (validation surfacing, 7), C3 (composition & authoring, 6), C4 (run
 observability, 8), plus the six C5 entries above — 27 entries. The method is