feat(ai): close four session-reliability gaps for retained provider sessions

[Dastari]

Summary

Closes four session-reliability gaps reported by a downstream consumer of the
retained Codex app-server deployment. Two of the four reported premises turned
out to be wrong at the pinned revision; this describes what was actually
broken.

Conversation bootstrap failed structurally during streaming. The snapshot
retry predicate compared row_version and stream_head, both of which every
coalesced live delta advances at roughly the coalescer rate. The bounded
snapshot therefore returned Conflict for exactly the sessions a user is most
likely to open, which a client cannot distinguish from a disconnection. The
predicate now covers only fields the bootstrap returns; a live delta appends a
session event and changes nothing in the payload. The watermark is documented
as a resume floor rather than an equality point.

Session-event streams ended as silence. RecvError::Closed returned with
no terminal envelope, so "stream over, resubscribe" looked identical to network
silence. A typed AiSessionStreamClose now accompanies every server-side end,
yielded before the existing error so current clients keep working.
Reauthorization distinguishes an unavailable dependency (bounded, per-session
jittered grace) from an authoritative denial (immediate), so one authorization
restart no longer drops every open stream simultaneously. A periodic bounded
durable head check makes single-replica delivery independent of the
process-local wakeup channel.

Failed runs carried no classification. The run_failed and
run_recovery_required events already existed and were already emitted in the
terminal write's transaction — the report that they were missing was stale. What
they lacked was any failure record. The payload advances to a v2 tagged shape
carrying a stable code, a retryable flag, and an admission reason, with the
reader accepting v1 and v2 and failing closed otherwise. retryAiRun authors a
new run over the same durable user message under current policy where
re-execution is provably safe; acknowledgeAiRunFailure dismisses a failure
without deleting audit history.

Retained provider threads were dropped silently. Ordinary user behaviour —
pressing Stop, a changed rule fingerprint, an exceeded budget — invalidated the
retained thread, and the next message started a fresh one while the durable
transcript still rendered as continuous. Every invalidation funnel and every
explicit rebind now appends a durable session event carrying only the
server-owned reason class.

Messages accepted during cleanup already converged across a restart — the
deferred turn is scheduled as a durable retry and claim_next reclaims
retry-scheduled runs, so delivery never depended on the Deferred outcome
reaching the executor. The gap was at the end of that allowance: exhausting it
propagated a conflict and left the run to expire into RecoveryRequired, which
is misclassified because nothing had executed, and stuck until an operator
intervened. It now closes as a clean, retryable, visible failure.

Interrupt handling reports what it proved via AiRunInterruptSettlement, which
fails closed. No adapter reports Settled: the Codex app-server
turn/interrupt response is an empty object, TurnStatus has a first-class
interrupted value, and a resumed thread pages prior turns back through
thread/turns/list, so an acknowledgement cannot distinguish a discarded
partial turn from a retained one. Treating it as settlement would let the model
carry content the durable transcript never recorded.

Verification

Local command output, run on this branch:

• cargo fmt --all -- --check
• cargo test -p graphql-orm, -p graphql-orm-storage,
  -p graphql-orm-backup --features orm-sqlite,
  -p graphql-orm-operation-catalog -p graphql-orm-ai-tool-profiles
• scripts/check-ai-provider-lanes.sh test
• cargo test -p graphql-orm-ai --features graphql-case-pascal --test graphql_naming
• cargo clippy -p graphql-orm-ai --all-targets across all five provider
  features plus local-harness, warnings denied
• cargo check -p graphql-orm-ai for postgres, mssql, and the mixed
  graphql-orm-ai/sqlite,graphql-orm/mssql lane
• warnings-denied, missing-docs-denied rustdoc for graphql-orm-ai,
  graphql-orm-operation-catalog, graphql-orm-ai-tool-profiles
• python3 scripts/check-documentation.py,
  python3 scripts/generate-workspace-inventory.py --check,
  python3 scripts/check-release-state.py,
  scripts/check-workspace-dependencies.sh,
  scripts/check-package-release-policy.sh <merge-base>,
  scripts/check-semver.sh <merge-base>,
  scripts/check-release-manifest.sh

New regression coverage:

• bootstrap under concurrent session-event churn, verified to reproduce the
  original Conflict against the previous predicate and to pass against the
  new one, plus a unit test pinning exactly which fields admit or reject a
  snapshot;
• stream close envelope on authoritative denial;
• reauthorization blip survival with retries proven to have occurred, then
  fast-fail on denial;
• durable replay fallback delivering with no in-process wakeup, using two
  Database handles over one SQLite file so their wakeup channels are genuinely
  independent;
• retry idempotency, refusal for RecoveryRequired and for an unclassified
  failure, acknowledgement after a refused retry, and revoked-principal denial;
• failed-run event and failure record replaying after a process restart;
• retained-thread invalidation disclosed on the sequenced stream and replayed
  from an earlier watermark;
• cleanup-pending deferral converging through a durable retry, and converging
  to a clean visible failure once the allowance is exhausted.

Documentation impact

Check exactly one option and explain the result.

• Documentation updated
• No documentation impact

Explanation: graphql-orm-ai 0.82.0 with schema module 0.61.0. Adds
crates/graphql-orm-ai/docs/session-reliability-adoption.md as the adoption
contract — exact new and changed public APIs, watermark resume-floor semantics,
retry admission rules, and the behaviours that changed with no API change —
linked from the component docs index. Updates the component CHANGELOG,
MIGRATION (schema module step, source-breaking changes, additive GraphQL
surface, new event types, v1-to-v2 payload), README, the conversation-bootstrap
section of capability-discovery-and-execution.md, and the generated workspace
inventory. Also states explicitly what multi-replica delivery would additionally
require, which this change deliberately does not supply.

[Dastari]

Added 4efa91d, a CI fix unrelated to the feature work but blocking this PR.

cargo-semver-checks 0.50.0 now fails outright when the only selected crate is a proc-macro crate, instead of silently checking nothing:

error: no crates with library targets selected, nothing to semver-check
note: skipped the following crates since they have no library target: graphql-orm-macros

That blocked every pull request regardless of content — this branch touches zero files under crates/graphql-orm-macros. scripts/check-semver.sh has always excluded the crate for the same reason, with the rationale in a comment; the CI job simply never matched it. Macro compatibility remains covered by the aligned package-version gate and the compile/trybuild fixture matrix.