[ako]

Syncs ako/mxcli:main into upstream: 12 commits from fork PR ako#1528. It fixes the flow-graph-geometry cluster: canvas positions, sides and merges that mxcli wrote differently from what a Studio Pro author would draw, and the describe and lint warnings around them.

Each defect has its own commit with a failing-first test, a control run, a mdl-examples/bug-tests/flow-geometry-*.mdl script and a finding.

Fixes #884, #923, #990 and #1587. Fixes cases 1–5 of #1598; case 6 is open, see "Not done here".

Fixes

Issue	Defect	Fix
#1598 case 2	Every note line was stored 2 → 0, so a note below its activity drew its line out of its own bottom	The sides now face each other, vertical first. Notes use the activity numbering, as Studio Pro's own documents do
#1598 case 4	The merge after a loop inside an IF landed inside the loop box	The layout now counts the loop's 100px iterator column
#923	while true with early returns gave three "NOT equivalent" warnings and three MDL-FLOW01 hits for a lossless flow	A branch walk stops at its own split
#990	The @anchor help said "each end of the outgoing flow", but to: is the flow arriving	Help, example, AST comments and skill are corrected, with the two-statement recipe for a vertical edge
#923	describe walked through a one-input merge (join finish; merge finish;) and warned that re-exec deletes it, which is no longer true	On the main line it is described as join junctionN; merge junctionN;. Elsewhere the warning now says what is true
#884	The ON ERROR rejoin merge was stored on the End event (560;200 over 560;200), or right of the next activity	The merge sits on the edge it splits. @merge still wins
#1598 case 3	The handler row sat 90px under its activity, inside the output label, and a THEN handler overlapped the ELSE row	ErrorHandlerPitch is now 130, Studio Pro's median, and the layout counts the handler row
#1598 case 5	Parameters were 100px apart whatever their names	The pitch follows the widest name, with a floor of 100, so names under 15 characters are unchanged. The old row still reads as derived
#1587	CASE outcome sides followed WHEN order only. @anchor(A: (from: right)) was refused (MDL092), and from: on a CASE was ignored	New syntax @anchor(A: (from: right), (empty): (…)) sets sides per outcome. CASE from: is the merge exit, describe round-trips only non-default sides, and MDL092 reports bad keys
#884	An IF/ELSE ending a loop body got a merge with no way out	No merge at a loop or while tail. @merge keeps one, and describe states a stored one
#1598 case 1	A note-only splice stored the note at x 3390 and shifted the row by +1180px	Notes take no room in a splice, and a stated note is pinned

The @anchor per-outcome form is new MDL syntax. It covers grammar, visitor, builder, describe, mxcli syntax microflow.layout, MDL_QUICK_REFERENCE.md, the docs site, the skill, and the doctype example (02c, pattern 7). The generated parser is not committed.

Evidence

The conventions were measured, not assumed, from Studio Pro-authored documents:

• Note lines: in 523 distinct note lines, the element side follows "vertical then horizontal" 98% of the time. Notes level with their activity store 1 → 3 / 3 → 1, which only the activity numbering explains.
• Handler rows: across 160 of them, the median is 130px below the activity (quartiles 125–150).
• Parameter rows: across 95 of them, the pitch grows with name length (160px median for 20–24 characters).
• Enum outcomes: across 1,374 of them, sides follow geometry. That isn't the default for three or fewer cases, because describe reads WHEN order from those sides.

Controls. Every test fails with the reported symptom when its fix is reverted. End-to-end on an 11.13 app, comparing main and this branch:

	main	this branch
loop-in-branch merge	1070;200 (inside the box 790..1110)	1170;200
ON ERROR rejoin merge	560;200, the End event	425;200
note-only splice	$E 2410→3590, note 3390;600	nothing moves, note 1500;600
describe warnings for the two while true test microflows	3 / 1	0 / 0

mx check: app copies with every new bug-test script and the full 02c doctype file give the same error set as the pristine copy.

Studio Pro 11.15.0-rc.4, via MCP: ped_check_errors reports nothing for the loop merge, the parameter rows, the note lines, the CASE flows or the loop-tail flow. Studio Pro stores exactly the positions this branch writes.

make build, make test, make lint, make check-mdl and make check-findings pass, and docs-wiki/bug-patterns/flow-graph-geometry.md is synced.

Not done here

• Microflow canvas layout: six cases where mxcli writes positions, sides or splices that Studio Pro would draw differently or overlap #1598 case 6: an if/else inserted in a splice shifts the decision, merge and return by +339, but not the replaced activities. A re-run is then refused, and the no-else variant stacks two decisions. This needs a behaviour choice, so it is proposed as its own task:
  • pin stated positions and refuse a shift past them, or
  • carry the shift into later fragments.
• The MCP backend's create path sends no connection sides and no notes, so @anchor has no effect over --mcp. This is a separate follow-up.
• Known gaps: a CASE at a loop tail still gets a merge with no way out, and @merge on CALL WEB SERVICE is still ignored