[viethqb]

Summary

Workflow parity roadmap package #2a, "End hooks, Part A": finally: now runs on every flowable and on the flow, and Loop, LoopUntil and WorkingDirectory run their own errors: and finally:, following Kestra's list rule (tasks, then errors only on failure, then finally; a block's and the execution's state come from the main tasks only, so a failing finally changes nothing). afterExecution: and the Rule 2 sweep fix are #2b. Pause finally: moves to #9 and is refused naming it.

Closes #

Changes

• Model reads flowable finally: and the loops' errors:; the plan appends the new lists after today's layout, so every stored task_runs.task_ix stays valid (a frozen layout fixture and a prefix test guard it).
• One list rule (current_list) in the executor; advance stays pure. Parallel/Dag/If/Switch/AllowFailure/Sequential/LoopUntil/WorkingDirectory and the flow use it. A Dag with errors: keeps its dependsOn gate (it used to ignore it).
• A Loop iteration runs the Loop's own errors/finally. A WorkingDirectory's hooks run after the block, outside the directory (no block id, no snapshot access).
• Worker: one shared spec::path::find_task replaces five walkers that missed lists. This also fixes two pre-existing gaps (a Loop or WorkingDirectory inside an errors: list).
• A stored execution whose flow has a Pause with finally: still plans as a Pause; publish and new starts refuse it naming Phase 3b-1 — Query Engine (async execution, REST polling) #9.
• Oracle ratchet 301/72/72/0 to 301/60/60/0: 12 flows bank. task-runif-taskswithstate stays baselined (renderer empty test, package Phase 0b — Frontend skeleton (dual-portal SPA, sparklabx design system) #5).
• Docs: spec and plan for Phase 1a-1 — AuthN & Session (local login, sessions, CSRF, bootstrap) #2 (EN), dated notes in the roadmap, spec Phase 0a — backend bootstrap & dev infrastructure #1 §10 and the programme spec (EN and VI).
• No migration.

Test plan

• Full cargo test -p egret-workflow --no-fail-fast on a private Postgres/Valkey before the final rebase: only the 7 sql_starrocks_live tests fail (need a StarRocks engine, as on main). After the rebase onto fix(catalog): classifications follow a renamed table or column #757: lib 1821 pass, executor ratchet 61/61 pass.
• Each task had a spec and quality review; the whole branch had an adversarial review whose Important finding (stored Pause with finally) is fixed and pinned.
• Mutation checks recorded per rule.

Deploy note

Stop old workers before deploying (a re-plan must not interleave with an old binary). Two edge cases are documented, not fixed: a malformed finally: stored before this PR stalls an in-flight execution of that flow until it is killed, and a task id duplicated between a pre-existing finally: and another task resolves to the first match for executions already running.

Documentation

Spec, plan and roadmap notes are in this PR. The VI mirror of the new spec and plan lands with #2b (task B8). The platform user docs live in egret-data/docs: a follow-up is needed (a finally: now runs; Pause finally: is refused).

Checklist

• Linked to an issue
• Tests added or updated
• Docs updated in egret-data/docs (follow-up)
• Commit messages follow Conventional Commits
• No secrets, tokens or real data added

🤖 Generated with Claude Code