GeezerVison

…rong (#182)

Closes #180. Based on main (#178 and #174 are merged there); not stacked
on #172 (see below).

## What changes, per proposal

When a call needs the states of several state sets (Door: `d.lock();
d.open();`, `open` needs `doorState(this) == closed && lockState(this)
== unlocked`):

- **Failure marked where it is.** Which set is wrong comes from the
server's `errorContext` per diagram, not from message text: a set whose
`invalidStates` is not empty is in a wrong state; a set whose possible
states are all accepted, with another set in a wrong state, is "right".
In a right set, a transition of the call from the actual state whose
condition names the wrong set is drawn as usual (white, normal arrow)
and only the clause the actual state does not meet is red, also when
conditions are collapsed. In the wrong set the call is red as before.
Every other transition of the call stays red as now: when it has no
condition on the wrong set, when the condition is not only states, when
no set is wrong on its own (the verifier can only reject the
combination), or when the state is unknown.
- **Wrong set first.** `orderStateMachines` puts the sets in a wrong
state first, then the other sets of the error, then the rest. Focus mode
fits the first diagram as before, so the ✕ state is now on screen
(A1_38: `mark` first, the case from the IDE review).
- **Self-loops.** A transition that keeps its state is drawn as a round
loop (same ends, same side, as far out as before so it stays inside the
laid-out SVG, slightly thicker) instead of ELK's flat curl the size of
its arrowhead. It is redrawn after Mermaid renders, from a `%%
lj-self-loops:` comment like the existing dimmed-transitions one.
- **Conditions as in the spec, before and after the method.** A label
reads `[condition before] method [condition after]` on one line, e.g.
`[lockState(this) == unlocked] open`. Each condition is the spec text,
so it reads like the Expected box and the spec, and a condition with
other clauses stays readable: in `[lockState(this) == unlocked &&
size(this) > 0] open` only the clause the actual state does not meet is
red. Edge labels no longer wrap at Mermaid's 200px width, so a condition
stays on the line of its method.
- **Set names in the summary.** "expected doorState `closed`, lockState
`unlocked` · actual lockState `locked`": expected in declaration order;
actual lists only the sets in a wrong state (all known ones if none is).
A set in which every state allows the call is left out of both sides,
for named and unnamed sets (the A1_56 "(undoable or notUndoable) and
redoable" case). Unnamed sets keep the current wording.

Single-set classes are unchanged except for the round self-loops and the
summary naming a named set (`expected phase open · actual phase
closed`).

## Ways to fix and #172

Not in this PR. Highlighting `unlock` on the diagram needs to know which
calls a hint suggests, and `HintDTO` only has `kind` and `message`;
getting the calls from the message text, or from paths in the diagram
(judged unsound in #171), are both out. #172 lists the verifier's hints
in a "Ways to fix" section and does not draw them; drawing a hint is
#156, which needs a structured hint from the verifier first. So this PR
does not depend on #172 and does not touch the hints; the two only meet
in `fsm.ts`/`script.ts` call sites and should rebase cleanly in either
order.

## Screenshots

Headless VS Code, this branch's server (verifier 0.0.37, fsm 0.2.3).
Focus mode is what opens from the error.

**Door** (the spec from #178, `d.lock(); d.open();`):

| | Before (main) | After |
|---|---|---|
| Focus on error | <img
src="https://raw.githubusercontent.com/liquid-java/vscode-liquidjava/pr-screenshots/180/door-focus-before.png"
width="380"> | <img
src="https://raw.githubusercontent.com/liquid-java/vscode-liquidjava/pr-screenshots/180/door-focus-after.png"
width="380"> |
| Show all | <img
src="https://raw.githubusercontent.com/liquid-java/vscode-liquidjava/pr-screenshots/180/door-all-before.png"
width="380"> | <img
src="https://raw.githubusercontent.com/liquid-java/vscode-liquidjava/pr-screenshots/180/door-all-after.png"
width="380"> |

**Study exercise A1_38** (BufferedInputStream, sets `stream` and `mark`;
`buffered.reset()` without `mark()`; proposal-D-task-types with
liquidjava-api 0.0.8):

| | Before (main) | After |
|---|---|---|
| Focus on error | <img
src="https://raw.githubusercontent.com/liquid-java/vscode-liquidjava/pr-screenshots/180/a138-focus-before.png"
width="380"> | <img
src="https://raw.githubusercontent.com/liquid-java/vscode-liquidjava/pr-screenshots/180/a138-focus-after.png"
width="380"> |
| Show all | <img
src="https://raw.githubusercontent.com/liquid-java/vscode-liquidjava/pr-screenshots/180/a138-all-before.png"
width="380"> | <img
src="https://raw.githubusercontent.com/liquid-java/vscode-liquidjava/pr-screenshots/180/a138-all-after.png"
width="380"> |

The right set, scrolled down in focus mode (call drawn as usual, unmet
condition red). These two and the A1_38 shots are from before the change
to conditions as in the spec, which they still show as `[lockState:
unlocked]`; the Door shots above are current:

| Door `doorState` | A1_38 `stream` |
|---|---|
| <img
src="https://raw.githubusercontent.com/liquid-java/vscode-liquidjava/pr-screenshots/180/door-focus-second-after.png"
width="380"> | <img
src="https://raw.githubusercontent.com/liquid-java/vscode-liquidjava/pr-screenshots/180/a138-focus-second-after.png"
width="380"> |

## Not in this PR

- Marking states that are current and fine (`closed` in `doorState`,
`open` in `stream`), as agreed.
- Highlighting the ways to fix on the diagram (above).

## Reviews

- Codex: 5 rounds (comments below); rounds 1-4 found one or two issues
each, all fixed; round 5 found nothing.
- Clean-code audit: comment below; one fix, the rest declined with
reasons.

## Tests

- client: `npm run test:unit` 128 passed (5 new, each fails without the
change: order, call/condition marks with `!=` and collapsed conditions,
round loop, named summary without the any-state set; one existing
expectation gains the self-loop comment), `npm run test:typecheck` and
`npm run lint` clean.
- server: no change; `mvn -B test` 39 passed.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <[email protected]>