fix(dag): session-following location stamps + adoption-vs-deletion fencing (#269, #270) - #272

Root cause (#270, shared territory with #269): admission was check-then-act —
an instance re-read the durable row (ownsWorkflow) and then published its
in-memory entry / materialized a child session. A deletion (Session.remove → FK
cascade) that commits between the passed ownership check and the action slipped
past it: the in-flight spawn could create a child for a workflow the cascade had
already removed. SpawnReady eviction was the mitigation, not a fence.

Fix (Model (b) conditional-claim fence of the accepted atomic-adoption design —
status+directory conditionality, no lease column):
- C2: DagStore.tryClaimAdoption — a single conditional UPDATE that matches the
  row only while it STILL EXISTS and is non-terminal, returning whether the
  admission claimed it. It converts the admission from check-then-act into
  act-with-condition; a committed deletion/terminal matches zero rows and the
  admission aborts atomically. Directory ownership is deliberately NOT re-done
  here: ownsWorkflow canonicalizes both sides and a SQL directory comparison
  would diverge (create stamps are realpathed, Moved re-stamps are not).
- C3: recoverWorkflow and the WorkflowStarted handler gate runtimes.set on the
  claim (after the ownsWorkflow directory revalidation), so a cascade in the
  final window aborts the adoption before it ever publishes an entry.
- C4: spawnNode re-admits atomically right before sessions.create — the fence
  that closes the spawn window nodeQueued's guard alone leaves open between its
  read and its publish. A committed deletion matches zero rows and the spawn
  aborts before any child session exists.

Probe C7 (red-first, park-gate style): parks the window-2 claim, commits a
deletion, releases, and asserts no post-deletion child survives.
RED evidence (fence conditionality reverted → always-claim): a child session is
created-then-cancelled for the deleted workflow, so `expect(cancelsA).toEqual([])`
fails (Received +3). Restoring the conditional claim makes the claim return false
on the deleted row and abort before sessions.create (GREEN).

All existing DAG-LOC-01 guards stay green (24 probes).

…ext(a) (#269)

Root cause (#269): the atomic-adoption redesign narrows the DAG-LOC-01
write-once directory invariant into a two-writer whitelist. R7-ext(a) previously
asserted "no UPDATE writes the directory column anywhere in the dag trees," which
held but left the new whitelisted writer (the Moved re-stamp) unpinned.

This is an invariant UPDATE, not a weakening — mandated by #269's own acceptance
("the directory stamp must move WITH the session in one transaction"):
- The ONLY sanctioned directory writers are the dag projector's WorkflowCreated
  INSERT (create-time stamp, onConflictDoNothing) and the session projector's
  SessionEvent.Moved re-stamp (payload-sourced, no SessionTable read).
- R7-ext(a) keeps its negative half (no .set({directory}) in the dag trees) and
  gains a positive whitelist half asserting the Moved re-stamp exists in the
  session projector (outside the dag trees, so R7/R7-ext stay scoped correctly),
  pinning the #269 resolution at the static-contract level.
- WorkflowTable.directory column comment documents the two-writer whitelist.

RED evidence (mutation reversion — Moved re-stamp removed from the session
projector): the whitelist expect fails, Expected true / Received false.

No user-visible semantic change; no runtime behavior altered in this commit.