fix(agileplace): clear planned dates with RFC-6902 remove, not null replace

…ecks retired

The second 2026-07-21 smoke run (after the issue #52 remove-op fix merged)
passed the full sequence: step 6 read back isBlocked=False and both planned
dates None, confirming the RFC-6902 remove clear and the blocked-state clear
live. Every AgilePlace [live-check] item now has a recorded live outcome.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The io v2 public docs do not document the Dependencies feature's
endpoints, and nothing gets built here against guessed shapes (issue #52
made the case). probe_dependencies.py answers the read-surface question
against a real tenant: dump a card's GET keys (dependencies may already
be embedded), verify plumbing via the documented connection/children
control endpoint, then GET six candidate dependency paths and report
status + body for each, with a devtools-capture fallback instruction
when everything misses.

Strictly read-only -- every request is a GET, and the test suite pins
that invariant. Ships with the agreed design spec for the full
blocked-by -> Dependencies feature (phases 0-2).

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Three confirmed P2s from the local Codex (GPT-5.5, xhigh) pass:

- The missing-env test deleted environment variables that env_config()
  immediately restored from .env on a configured checkout -- exactly the
  machine the probe is meant to run from. The test now points
  config.ENV_FILE at a nonexistent path.
- Candidate classification conflated inconclusive with missing: a 204
  (route exists, empty collection) was neither found nor missed, and
  429/5xx counted toward the all-missing devtools fallback. Now 2xx =
  found, 404 = miss, anything else = inconclusive, and the definitive
  fallback only fires when every candidate returned a clean 404.
- URLError/TimeoutError escaped probe_get as a traceback mid-diagnostic;
  transport failures now return status 0 and classify as inconclusive.

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

First live probe run (production tenant) confirmed the undocumented read
surface: GET /io/card/{cardId}/dependency answers 200 with
{"dependencies":[]} -- singular path, matching the connection/children
pattern. Five sibling candidates 404'd cleanly against a passing control,
and the card GET embeds no dependency data. Populated entry shape and the
write endpoint/body remain [live-check] pending (Phase 0b).

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The UI-created EP-3A -> JPOWER1 dependency (mirroring GitHub edge #31 ->
#68) read back as direction/cardId/timing/createdOn with timing
'finishToStart' -- and no dependency id, implying pair-addressed deletion.
Write and delete request shapes remain [live-check] pending on a devtools
capture.

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

DELETE /io/card/dependency with {cardIds:[dependent], dependsOnCardIds:
[blocker]} -- pair-addressed as the id-less read entries implied, plural
batch arrays in the card/connections style. Create shape remains the one
pending [live-check].

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…confirmed

POST /io/card/dependency with the delete's pair body plus explicit
timing:'finishToStart'. The POST response embeds a face projection with
dependencyStats (native health counters). Read/create/delete are now all
live-confirmed; only duplicate-create behavior remains fact-finding for
the smoke round-trip.

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…dencies

Phase 1 of issue #57, built entirely against the live-confirmed shapes
(API-VALIDATION.md 'Dependencies API discovery'):

- agileplace.py: card_dependencies (fail-closed read -- None means
  unknown, never empty), incoming_dependency_ids, and batch
  create/delete_dependencies using the captured pair bodies with
  timing finishToStart.
- sync.py: sync_dependencies reconciles after the Blocked-flag pass,
  guarded by the same complete-snapshot check. Every GitHub edge is
  mirrored (Done blockers included -- the edge is structural; native
  dependencyStats display satisfaction), GitHub is authoritative only
  between two sync-managed cards, failed reads skip the card entirely
  (duplicate-create is unconfirmed live), plan-only ids never cross a
  read boundary, and timing is human-owned after creation (pair-matched
  reconciliation).
- smoke.py: steps 11-12 add the dependency create/read/delete round-trip
  plus duplicate-create fact-finding (INFO); stale probe renumbered 13.
- Offline doubles in test_run.py and test_smoke.py serve the confirmed
  endpoint shapes; the dry/apply parity harness normalizes the new
  depend/undepend mutations.

The Blocked flag is untouched (phased decision -- spec Phase 2).

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

test_run._configure delenv'd GH_PROJECT_* but env_config() re-reads .env
on every call and repopulates them, so on a configured checkout the
offline world reached for the developer's REAL GitHub Project and failed
with 'unexpected gh command: project item-list ...'. Reproduced by
simulating a configured checkout, fixed by pointing config.ENV_FILE at a
nonexistent path (the pattern test_config_env.py and test_smoke.py
already use), verified green with and without a real .env present.

Closes #59

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

… double

The 2026-07-21 smoke run (suffix 10a01e) retired the dependency round-trip
live-checks: create/read/delete all behaved as coded, and the duplicate-
create fact-find answered HTTP 409 Conflict 'Dependency already exists'
(not idempotent). The offline FakeTenant now raises that exact rejection,
which also exercises smoke's duplicate-probe SystemExit path offline, and
API-VALIDATION.md records the outcomes -- no dependency [live-check]
remains open.

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…al review

Three confirmed findings:

- P1: an active issue blocked by a retired (NOT_PLANNED/DUPLICATE) issue
  with a URL-owned card lost its edge from the desired set (blockers
  resolved through active issues only) while the retired card stayed in
  the managed set -- the valid native dependency would be deleted as
  stale on every run. Blockers now resolve through _blocker_cards, which
  covers retired issues via retired_card_by_url, mirroring how the
  Blocked-flag path already treats retired issues as Done blockers.
- P2: card_dependencies accepted entries without a usable direction or
  cardId and would treat a paginating response as a complete snapshot;
  acting on such a read re-creates an existing pair and the live 409
  aborts the run. Reads now fail closed on any pageMeta member, unknown
  direction, or missing cardId.
- P2: the main-level harnesses (test_sync_main, test_retired_issues,
  test_hierarchy_ownership) mocked every AgilePlace boundary except the
  new dependency read, so the suite made real DNS lookups silently
  converted to None -- the source of the flaky 3s-vs-58s suite times.
  All three now stub card_dependencies; the suite runs in ~0.3s flat.

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ans now

Issue #57 Phase 2, per maintainer decision after judging the native
depends-on card icon on the live board:

- sync.py sheds ALL Blocked-flag code: the flag-text mirror in step 4,
  the flag-clear during card retirement, and the now-dead
  stages.blocked_reason. The sync never writes /isBlocked or
  /blockReason; a human's flag (any reason text) is never touched.
- clear_legacy_blocked_flags.py: one-shot throwaway batch-clear for the
  flags the old behavior left on the board. Signature-matched
  (fullmatch on 'Blocked by #N[, #M...]') so human-authored flags
  survive; dry-run by default; idempotent; to be DELETED with its tests
  once the board is clean.
- ops_blocked/card_is_blocked/card_block_reason stay in agileplace.py as
  validated API surface (smoke still exercises the shape; the script is
  their remaining writer).
- Tests pin the new contract: retirement is lane-only, no flag scrubbing,
  and no PATCH at all for a card already in its Done lane.

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

clear_legacy_blocked_flags.py ran against the production board on
2026-07-21 (34 sync-authored flags cleared, 0 human flags touched) and
was always destined for deletion afterward -- it lives on in git history.

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Codex (gpt-5.6-sol, context-primed pass) flagged that README steps 1/4,
the PATCH-batching paragraph, the module layout, sync.py's module and
_retire_card docstrings, and smoke's step 5-6 docstring still described
the retired flag mirror. All now state the Phase 2 contract: blocked-by
mirrors as native dependencies, the Blocked flag is human-owned, and
smoke's blocked round-trip validates API surface only.

Part of #57

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The live create response is sparse: it carries the new card id but no
version and no customId/laneId echo (confirmed during the first live
migration, 2026-07-21). Indexing it as the card's snapshot queued
redundant /customId and /laneId ops, which the issue-#8 stale-ops guard
then read as a concurrent edit -- every fresh create+sync --apply
aborted before any metadata landed and no state was saved.

After an apply-mode create that returns an id, refetch the card once and
index the fresh card as the snapshot. No redundant identity ops are
queued, and the fresh snapshot's usable version lets the batched PATCH
skip _card_with_version's own refetch, so net API calls are unchanged.
A failed or mismatched refetch falls back to the sparse response rather
than failing a run whose create already succeeded. Dry runs are
untouched: planned-card snapshots flow exactly as before.

The offline sync-run double now mirrors the sparse create response
(id-only) and serves the persisted card via the single-card GET, pinning
this path offline; API-VALIDATION.md records the new live fact.

Closes #55

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Adversarial review of the issue-#55 fix flagged that only the sync-run
double was aligned with the live-validated create contract: the smoke
FakeTenant still echoed the request body plus a default version -- a
shape the live tenant never returns (version absent confirmed
2026-07-20/21; no customId/laneId echo confirmed 2026-07-21). Any future
smoke code reading an echoed field would stay green offline and break
live -- the same masking pattern that let #55 reach the first migration.

The double now returns id-only, unconditionally, while still persisting
the full card for the single-card GET to serve back. The
create_returns_version knob is gone: a sparse default left it unused,
and the versionless-create test now pins the default contract directly.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>