showcase: 4/5 projects and 9/10 tasks are silently rejected on every boot (FSM initialStates vs seed data)

[os-zhuang]

Summary

The showcase seeds 4 of 5 projects and 9 of 10 tasks are silently rejected on every boot. The running app has exactly one project and one task. Nothing in the boot log says so.

The platform is behaving correctly — this is a self-contradictory fixture plus a surfacing gap.

Root cause

showcase_project declares its own FSM entry gate (#3165):

// project.object.ts
events: ['insert', 'update'],
initialStates: ['planned'], // "without it a project could be born already completed"

The project seed in the same repo creates 5 projects with status:

Project	seeded status	legal on insert?
Website Relaunch	active	✗ rejected
Data Platform	active	✗ rejected
Compliance Audit	on_hold	✗ rejected
Mobile App	planned	✓ survives
Legacy Sunset	completed	✗ rejected

Seeds deliberately run validation — SeedLoaderService.SEED_OPTIONS = { isSystem: true, skipTriggers: true }, commented "Lifecycle HOOKS (derived/default fields, validation) still run." So the gate fires and correctly rejects the four.

showcase_task.project is a master-detail to project, and 9 of the 10 seeded tasks point at the three projects that were never created. Only App wireframes → Mobile App survives.

showcase_project_membership (references team + project) lands 0 of 3 for the same reason.

Evidence

Built artifact is correct — the loss is at load time:

dist/objectstack.json: showcase_project records=5 showcase_task records=10
sqlite dev.db: showcase_project 1 row showcase_task 1 row
 showcase_project_membership 0 rows (3 declared)

knex debug on a wiped DB shows exactly one single-row insert per table:

insert into `showcase_project` (…) values (?, …) returning * ← 1 statement
insert into `showcase_task` (…) values (?, …) returning * ← 1 statement
insert into `showcase_account` (…) select ? … union all select ? ← batched, 14 rows ✓

Impact

The showcase's own seed docstring says it is "sized to feed every view: every Kanban column is populated, tasks carry due/start/end/created dates (calendar, gantt, timeline) and a work location (map)". None of that is true at runtime:

• Chart Gallery — "One widget per chart family… honest visual coverage" — renders 14 charts on a single data point. Total Tasks 1, every bar/pie/funnel/radar/treemap/sankey degenerate, both month-bucketed trend lines a single point.
• Kanban: 1 card in 1 column (5 columns were intended).
• Gantt / timeline / calendar / map: one item each.
• Project list: 1 row of 5; no status or health variety, so the health-progression FSM demo has nothing to show either.
• This also blocks v16.0 browser test checklist — exercise every 16.0 feature in the running app #3358 §3 "Report drill-through range" — there is no meaningful date bucket to drill.

Silent failure

SeedLoader does record each rejection (this.logger.warn('[SeedLoader] …'), and buildResult carries the errors), but not one [SeedLoader] line reaches os dev output — I ran with OS_LOG_LEVEL=debug LOG_LEVEL=debug DEBUG=* (19k lines of knex tracing) and grepped case-insensitively: zero hits. The boot banner's only seed mention is seeded on empty DB · dev only.

So a fixture can lose 90% of its rows with no signal at all. Even if the fixture below is fixed, a boot-time summary of seed inserts/rejections would have caught this immediately — and would catch the next one.

Options for the fixture

1. Two-phase seed (respects and demonstrates the FSM). Insert all 5 as planned, then walk them to their target status via mode: 'update' datasets. The transition map allows planned→active and active→{on_hold, completed}, so it needs two update passes (a single pass cannot do the 2-hop jumps). Needs confirmation that ordering between multiple datasets on the same object is deterministic.
2. Seed everything as planned. One-line fix, but the project list loses all status/health variety and several demos flatten.
3. Widen initialStates. Cheapest, but it deletes the point of enhancement: state_machine 校验规则在 INSERT 上是 no-op —— 无法强制记录的初始状态(declared ≠ enforced) #3165 — "a project could be born already completed" is exactly what the gate exists to prevent.

I'd take (1) if same-object dataset ordering is guaranteed, otherwise (2) plus a follow-up. Flagging rather than picking, since it's a showcase-design call.

Found during the #3358 v16 browser sweep (§3).

[os-zhuang]

补充这个 bug 的用户可见影响(2026-07-23 dogfood,制作内部培训 PPT 时逐面截图,os dev --ui --seed-admin 全新库):

• All Views 看板整板空:唯一活下来的任务状态是「已完成」,不在看板四列(待规划/待办/进行中/评审中)中 → 四列全 0 卡,但底部计数「1 条记录」——观感是"丢数据"。列外记录静默消失这个前端侧问题单独记在 objectui#2792(即便种子修好,用户数据也会踩)。
• Chart Gallery / Command Center 全靠 1 条任务撑:柱图单柱、Total Spent 0、任务吞吐趋势单点——marquee 演示页的说服力直接归零。
• Projects 列表只剩 Mobile App 一行,甘特/日历同样近空。

两点建议顺手放这:

1. 修复时把种子状态分布对齐演示面一起做掉(看板各列有卡、甘特有跨度、图表有分布)——种子是为演示设计的数据,不只是"能插进去的数据";
2. showcase live smoke 加「看板至少 1 张卡」「KPI 非全 0」断言,这类"配置正确但观感崩坏"的回归,validate 门和单测天然抓不到。