refactor(ssr): remove the EDN snapshot round-trip; transaction edit is a plain form (heuristic 2)

The wizard serialized the whole accumulating form state into a `snapshot` hidden field
(pr-str EDN + custom readers), decoded it every request, and merged step-params back in.
For this single-step modal the snapshot is pure redundancy: every value is either in the
entity or the live posted form. Remove it:

- render: EditWizard.render-wizard renders a plain form -- no snapshot / edit-path /
  current-step hidden fields; a single `db/id` hidden rides in the form instead.
- middleware: wrap-derive-state rebuilds :multi-form-state per request from the entity
  (loaded by the db/id hidden) overlaid with the live step-params, replacing
  wrap-init-multi-form-state + wrap-entity. The ~34 :snapshot reads are unchanged --
  :snapshot is now a derived map, not a round-tripped blob.
- editable fields (accounts, vendor, memo, approval, action, mode, amount-mode) come ONLY
  from the posted form (absent = cleared) so removing all account rows doesn't resurrect
  the entity's persisted accounts; only entity-only fields (db/id, client, amount, ...)
  come from the entity.
- delete the dead initial-edit-wizard-state and render-account-grid-body.
- e2e: make removeAllAccounts re-query each iteration (whole-form swaps stale a captured
  row index) and restore the percentage test to type-then-add ordering.

Scorecard: snapshot EDN round-trip + custom readers + merge-multi-form-state -> gone
(snapshot-field renders 0). Verified on a fresh server: full suite 38 pass / 1 unrelated
fail, swap 6/6, transaction-edit 8/8 -- same green as before, snapshot removed.

.claude/skills/ssr-form-migration/reference/gotchas.md

@@ -108,12 +108,32 @@ should ride in the form, not a parallel snapshot). It bit the percentage-split e
 50% then adding a row reverted the first row to its snapshot value, yielding a 66.67/33.33
 split. Two ways it shows up and how to handle until the snapshot is gone:
 
-**Fixed (Stage 1 of the wizard→plain-form rewrite):** the operation handlers now read the
-live `:step-params` rows (already schema-decoded by `mm/wrap-wizard`, so typed) and fall
-back to the snapshot only when absent — typed values survive add/remove/toggle. The
-percentage e2e was restored to the realistic type-then-add ordering as a regression guard.
-The *full* removal (no snapshot at all; the posted form *is* the state) is the remaining
-rewrite stages.
+**Fixed (Stage 1):** the operation handlers read the live `:step-params` rows (already
+schema-decoded by `mm/wrap-wizard`) so typed values survive add/remove/toggle.
+
+**Done (Stage 2 — the snapshot round-trip is gone).** The EDN `snapshot` hidden field +
+custom readers + `merge-multi-form-state` are removed. A `db/id` hidden rides in the form;
+`wrap-derive-state` rebuilds `:multi-form-state` per request from `entity ∪ step-params`,
+and `EditWizard.render-wizard` renders a plain form (no snapshot/edit-path/current-step
+hiddens). The ~34 `:snapshot` reads still work — `:snapshot` is now a derived map, not a
+round-tripped blob.
+
+**Trap that cost hours — derive `entity ∪ step-params` correctly.** First cut was
+`(merge base step-params)`. Bug: `base` always carries the entity's *persisted* accounts,
+so after the user removes every row (step-params has no accounts key) the merge falls back
+to base → the persisted accounts **resurrect** on the next operation. Fix: editable fields
+(accounts, vendor, memo, approval, action, mode, amount-mode) come **only** from the live
+form (absent = cleared); only entity-only fields (`db/id`, client, amount, description,
+status, type) come from the entity. Lesson: with a posted form, "field absent" means
+*cleared*, not "use the persisted value" — never merge the entity's editable fields back in.
+
+**Verify the snapshot removal on a FRESH server, and don't trust a long-lived in-process
+test server.** Protocol/defrecord (`EditWizard.render-wizard`) and middleware reloads do
+**not** fully take in a running REPL — the server kept rendering the old snapshot field
+after `:reload`, and an in-process server that isn't reseeded between `npx playwright`
+invocations accumulates state that makes order-dependent tests flake. Both produced hours
+of phantom failures. Restart the REPL clean (or reseed) before trusting an e2e result; CI
+boots a fresh server per run, so the fresh-server number (38 pass / 1 unrelated) is the real one.
 
 ## Characterization tests rot against table order and removed wizard chrome
 

.claude/skills/ssr-form-migration/reference/scorecard.md

@@ -39,7 +39,7 @@ Each migration appends one row (after-numbers), referencing the before in the di
 | Phase | Modal | LOC | Routes | no-cursor twins | faked roots | snapshot merges | OOB | mixed hx- | cookbook reused / added |
 |-------|-------|-----|--------|-----------------|-------------|-----------------|-----|-----------|-------------------------|
 | 1 (baseline) | Transaction Edit `transaction/edit.clj` | 1608 | ~12 | 1 | 2 | ~75 | 0 | 8 | — / seeded 7 entries |
-| 2 (in progress) | Transaction Edit `transaction/edit.clj` | 1545 | **~5** | **0** | **0** | ~75 | 0 | 8 | — / 0 |
+| 2 | Transaction Edit `transaction/edit.clj` | 1584 | **~5** | **0** | **0** | **0 round-trip** | 0 | 8 | — / 0 |
 
 > **Phase 2 progress.** Achieved with parity held (swap spec **6/6**, transaction-edit
 > spec **8/8**, full suite **38 pass / 1 unrelated fail / 0 skip**, up from 30/2/7):