Round-4 review fixes: pipeline/count desync, persistence, CLI, analysis, export

[zntznt]

Fixes all twelve findings from the fourth multi-agent review, which targeted surfaces no previous round had covered: cli.js, the codegen/loops/assertions modules, persistence under stress, analysis numerics, scale, and round 3's own diff.

Each fix has a test that I verified fails against the unpatched code first. 251 unit tests, smoke green.

Findings fixed

#	Severity	Area	Fix
1	high	properties panel	+/- and Amount on a Delay or Queue move the pipeline, via SimEngine.setLiveCount
7	high	persistence	undo and redo write the autosave slot (_persistAutosave)
8	high	persistence	File > Open confirms first and commits the swap, so Ctrl+Z works
3	high	cli	--param, --steps, --runs, --pass-rate parse strictly and exit 1 on junk
5	high	codegen	eco.set() keeps the pool's resource type instead of greying it
6	medium	loops	depth limit 10 → 32; panel and CLI stop claiming "no loops" after a truncated search
9	medium	analysis	a batch saves and restores the live run's RNG position
10	medium	analysis	the sweep range re-seeds from the parameter actually picked
11	medium	export	PNG scale clamps to the canvas limit rather than saving a 0-byte file
12	low	analysis	Monte Carlo yields per step, not per whole trial
2	low	mobile	timeline drawer holds at 190px instead of 324px
4	low	cli	--help prints the header block, not implementation notes

Notable detail

Finding 1 is the panel-side sibling of the trader guard merged last round. The engine now refuses to let a trader pay out of a Delay or Queue, but the properties panel still called the raw model primitives on those same node types. A delay holding ten units in transit, minus three clicks, still delivered all ten: three units created from nothing. Three plus clicks stranded three forever in an empty queue.

Found while fixing, not raised by the review

• A delay never releases its authored stock unless you press Reset first. _queue was only ever built in reset(), but Run and Step bootstrap at step 0 with saveInitial() alone. Place a Delay, set Amount to 10, press Run, and the ten sit there forever. Same for a Queue. Both now seed through _seedPipelines(). This needs no mid-run interaction at all, so it is strictly worse than finding 1.
• Cancelling a Monte Carlo batch leaked the RNG. The driver dropped the generator instead of closing it, so its finally never ran. Harmless before the finding-9 change and corrupting after it; the driver now calls job.return().
• Two em dashes in the CLI header, which --help prints, so they were user-facing copy under the project's copy rule.

Measurements

• Monte Carlo on a 300-node model: worst event-loop gap 276ms → 22ms, for about 2% more total batch time.
• Mobile timeline drawer at 768x1024: 324px (32% of the screen) → 190px (19%), with the scrub row still inside it at the legend's two-row cap, tested at 768, 390 and 320 wide.
• detectLoops at depth 32: worst case ~12ms on a dense 120-node graph, bounded by the existing edge budget.
• --help: 56 lines → 45, with no implementation notes.

Testing

• node test/run.js — 251 passed (was 242 on main)
• npm run smoke — passed, no console or page errors
• Every new test baselined against the unpatched code: each fails without its fix.

One caveat

_pngScale's limits (16384px longest side, 2.0e8 total area) are conservative constants, not feature-detected. A browser with a lower ceiling than Safari's would still fail, though it now says so instead of silently saving an empty file.

---

Generated by Claude Code