migration 0059 over-sets is_blocked on dolt < 2.4.0 (dolt#11886) - stores that crossed it hide ready work

[seanmartinsmith]

hit this crossing my fleet from a main build at 1823f47 (schema 55) to 1.3.0 (f45b249): migration 0059 set is_blocked=1 on issues that aren't blocked, so they silently dropped out of bd ready. on the biggest db (1468 issues) ready went 334 -> 255

measured

bd 1.3.0 (f45b249), dolt sql-server 2.2.1, server mode, shared server with a dolt remote, windows 11. designated-migrator crossing (bd migrate schema --force), first on a sandbox copy with real data, then 27 live dbs:

• in dolt_diff_issues the 0059 commit is the only one touching is_blocked: 80 rows 0 -> 1 on that db
• e.g. an issue whose only edges are relates-to ends up is_blocked=1 and isn't in bd blocked
• bd recompute-blocked right after migrating: 322 rows corrected across 9 of 27 dbs (130, 96, 80, 5, 3, 3, 3, 1, 1). ready back to 334 on the big one
• bd doctor "Blocked State" flagged 65 on that db, the recompute corrected 80

cause

0059's recursive step (0059_recompute_null_gate_is_blocked.up.sql:250-257 at v1.3.0, unchanged on main 448c4b8):

JOIN dependencies d
  ON d.type = 'parent-child'
 AND ((r.kind = 'issue' AND d.depends_on_issue_id = r.id)
   OR (r.kind = 'wisp'  AND d.depends_on_wisp_id  = r.id))

type and both target columns are indexed, which is exactly dolthub/dolt#11886: the planner concatenates two index lookups and keeps only the OR as the residual filter, so d.type = 'parent-child' is dropped and blockedness propagates across every edge type. #11886's impact section quotes this join.

on the untouched schema-55 copy through the same 2.2.1 server, 0059's CTE as a SELECT returns 105. the same CTE with the recursive step rewritten without the OR (one join on d.depends_on_issue_id = r.id, kind and type in WHERE) returns 25, which matches the pre-upgrade build's own count. 105 - 25 = 80, the flip count.

ruled out: data (the old build's runtime count and the OR-free variant agree at 25), dirty state (clean dolt_status; only the 0059 commit flips rows). the runtime recompute (issueops/blocked_state.go, what bd recompute-blocked runs) doesn't use this shape and is correct on the same engine.

who's exposed

• fix is go-mysql-server#3899 (merged 09-18), first released in dolt v2.4.0 (09-30). every server before that has it: #11886 reproduced on 2.3.1/2.3.5, i measured 2.2.1. the ci pin in scripts/ci/install-dolt.sh:26 is 2.2.0
• bd's embedded engine: go.mod pins go-mysql-server at ...20260713210757-6d01d00bbbf3 (v1.3.0) and ...20260805191915-e5eafe0da809 (main, hotfix/1.3.1). both predate docs: list all integration pages in website nav #3899, so does fix(deps): bump dolt/go past the conjoin re-wording and re-open fixes (#6265) #6811's target (...20260915233322). i haven't reproduced embedded, that's inference from the pin
• same join in 0047_recompute_mixed_is_blocked, ignored/0007_recompute_wisp_is_blocked and ignored/0015_recompute_null_gate_wisp_is_blocked - the ignored twins run on every clone
• bd doctor only catches it in server mode, so embedded users get no signal

fix shape (0059's bytes are frozen, so not an edit to it)

1. now: a line in upgrading.md / the 1.3.x notes - after crossing 0059 on dolt < 2.4.0 or embedded, run bd recompute-blocked once per db on every clone. it's idempotent
2. a new main migration + ignored twin that recomputes with an OR-free recursive step (two UNIONed joins, which #11886 shows is correct), so stores that already crossed get repaired without anyone knowing to act
3. heads-up on Migration 0059 exceeds two hours on a 14k-wisp-dependency graph #6520: its proposed stand-in indexes would put the same shape on the wisp branch

detection: bd recompute-blocked --json - rows_corrected > 0 means you were hit

prior art: dolthub/dolt#11886 (fixed), #6520, #4138, #6608, #6716. nothing on the beads side tracks the stores that already crossed

[gascityinc-sarah]

Triage: kind/bug · priority/p1 — Migration 0059 hits dolt#11886, so blockedness propagates across every edge type: 322 rows wrong on 9 of 27 dbs, bd ready 334 to 255. recompute-blocked repairs it; embedded users get no signal.
Auto-triaged by @gascityinc-sarah. Reply or relabel if this is off.