feat(batching): IsolateBatchMembers error policy for opaque batch failures (GH-3289 Phase 3b)

[jeremydmiller]

Phase 3b of the #3289 batch-processing plan: item 3 (IsolateBatchMembers). (Phases 0/1/2/3a merged as #3299/#3300/#3301/#3302.)

The feature

Item 8 (ApplyItemException, #3302) handles failures the handler can name. This handles the opaque case — the batch handler throws but can't attribute it to a specific item. The new exception-type-keyed error policy isolates the culprit anyway:

// A deterministic error isolates the offending member...
opts.Policies.OnException<ValidationException>().IsolateBatchMembers();

// ...while a transient error still retries the whole batch (the policies compose by exception type).
opts.Policies.OnException<SqlException>().RetryWithCooldown(100.Milliseconds(), 1.Seconds());

IsolateBatchMembers() re-runs each member of the failed batch as its own size-1 batch (via the shared BatchReplay primitive from 3a), so only the member that actually reproduces the failure is dead-lettered while the healthy members succeed.

Implementation

• ProbeIndividuallyContinuation (+Source) — splits a batch (>1 member) into singletons; a size-≤1 batch is dead-lettered. That size-≤1 guard is the terminating/base case: a probed singleton that fails again just DLQs (no infinite re-probe) and it's also the "no effect on non-batched messages" path.
• IsolateBatchMembers() verb — added to IFailureActions / FailureActions / PolicyExpression, installed as an every-attempt ContinueWith. Composes with the retry verbs purely by exception-type matching (separate FailureRules), exactly as the plan's flagship example shows.

Tests (DB-free)

src/Testing/CoreTests/Acceptance/batch_isolate_members.cs:

• opaque-failure isolation — batch [good1, bad, good2] throws → probed into singletons → the two healthy items succeed and only bad is dead-lettered (captured via an IDeadLetterInterceptor, which runs under NullMessageStore);
• non-batched fallback — a solo message with the policy just dead-letters (nothing to isolate).

Both green; 253 existing error-handling/batch tests still pass. dotnet build wolverine.slnx -c Release clean.

Remaining plan follow-ups

• Item 2 — ProbeIndividuallyAfter(attempts: N): the count-based variant (probe only after N whole-batch failures) — needs a per-batch-chain threshold; next PR.
• CoalesceBy same-key poisoning for ApplyItemException (flagged in feat(batching): ApplyItemException to isolate poison items from a batch (GH-3289 Phase 3a) #3302): a small follow-up.

🤖 Generated with Claude Code