piker/ai/prompt-io/claude/20260610T173309Z_f15f8178_prompt_io.raw.md

NOTE: diff-ref mode entry; the AI output here is the committed doc itself (a design plan, not code) so no verbatim copy is duplicated:

git log -1 -p --follow -- ai/claude-code/plans/datad_service.md

Implementation deviation log

Where the landed 7-commit series differs from the plan as written (recorded for reviewer transparency; the plan doc is committed unmodified):

1. Commit ordering: the pytest config-dir isolation fix lands BEFORE the tractor-API drift port. The plan’s stage gates assumed a runnable+isolated test baseline; per-commit gating exposed that without isolation the paper-EMS test reads the user’s real (polluted) account.kraken.paper.toml and reddens.
2. The plan’s “stage 0: full pytest green” gate required first repairing pre-existing branch breakage vs tractor git main (boot AttributeError, stale discovery/exc/position APIs) — that repair became its own commit (“Port service+tests to latest tractor APIs”) rather than plan-stage work.
3. Stage-4 fail-fast placement: the plan said broker_init() raises on brokerd-ep-less backends; implementation moved the raise to spawn_brokerd() since piker ledger calls broker_init() directly even for paper accounts on datad-only backends.
4. brokers/_daemon.py change grouping: the import-cleanup hunks (exceptiongroup, _FeedsBus type-only import) landed with the caps-sec slim commit instead of the fixture-slim commit, keeping each intermediate tree import-clean without sub-hunk surgery.

Also of record: the plan’s “Verification” matrix was executed as written (per-suite gates each stage, the headless datad-feed smoke, the kucoin fail-fast unit check); the known pre-existing ~50% second-runtime-boot test wedge was characterized and excluded as a regression via revert-testing.