test(dispatch): unit tests for dispatch pipeline (was 575 LoC / 0 tests)

[Destynova2]

Summary

Adds 21 unit tests to src/server/dispatch/mod.rs, which previously had 575 LoC and zero tests — the largest blind spot in the test suite per the audit.

Tests are organized into the following groups within mod tests:

DLP phase (5 tests)

• dlp_clean_input_passes_through_engine — happy path: benign prompt does not trigger block.
• dlp_blocks_request_on_injection_match — edge: (?i)ignore previous instructions triggers DlpBlockError::InjectionBlocked.
• dlp_block_error_display_includes_pattern_name — protects the audit log entry built from block_err.to_string().
• dlp_disabled_scan_input_skips_engine — verifies the scan_input flag is honoured.
• dlp_no_engine_short_circuits_to_ok — Option::None DLP engine is the no-op precondition.

Cache phase (6 tests)

• cache_round_trip_returns_stored_response — happy path: put/get returns identical body.
• cache_miss_returns_none — empty cache lookup misses.
• cache_skips_oversized_entries — max_entry_bytes is enforced at insert time.
• cache_key_is_deterministic_for_same_request — identical requests hash to identical keys.
• cache_key_skipped_for_nonzero_temperature — temperature > 0 opts out of caching.
• cache_key_differs_per_tenant — multi-tenant isolation in cache keys.

DispatchResult variants (2 tests)

• dispatch_result_complete_carries_provider_and_model — non-streaming path preserves provider info.
• dispatch_result_fanout_carries_response_only — fan-out path stores only the response.

Audit entry construction (3 tests)

• audit_entry_dlp_block_uses_blocked_backend_marker — backend == "BLOCKED" on DLP block.
• audit_entry_provider_failure_uses_none_backend — backend == "NONE" on all-providers-failed.
• audit_entry_fanout_success_records_token_counts — fan-out audits both input and output tokens.

Provider behaviour (2 tests)

• mock_provider_returns_fixed_response — exercises the MockLlmProvider round trip.
• mock_provider_supports_only_configured_model — guards accidental model leakage in fallback chains.

Token accounting (2 tests)

• token_counts_propagate_from_provider_response — happy path: counters propagate intact.
• dlp_block_reports_zero_tokens_consumed — billing guard: blocked requests have token_counts: None.

Routing metadata (1 test)

• route_decision_default_route_type_no_complexity_tier — gate at Step 5.5 of the pipeline.

Honest scoping note

Happy path vs edge case (per task brief):

Category	Happy-path	Edge / negative
DLP	clean input passthrough, no-engine short-circuit	injection block, error formatting, scan_input=false
Cache	round trip, deterministic keys, tenant isolation	empty miss, oversized skip, non-zero temperature
Result	Complete, FanOut variants	—
Audit	DLP block, provider failure, fan-out success	—
Provider	mock round trip	unsupported-model guard
Tokens	propagation	DLP-block billing guard

Behaviours intentionally NOT covered (documented in the test module docstring as follow-up work):

1. End-to-end dispatch() happy path with the provider loop.
2. Provider fallback on 503 / rate-limit / circuit-breaker rejection.
3. dispatch_provider_loop() ordering with the adaptive scorer.
4. Audit log emission asserted from the AuditLog file (needs disk fixture).
5. Token-accounting double-count guard on retry.
6. Cache hit metric assertion (Prometheus recorder is global state — installing it twice in the same process panics).

The reason these aren't unit tests today is that every DispatchContext requires a fully built Arc<AppState>, and AppState has hard dependencies on:

• GrobStore (real disk + AES-256-GCM cipher init).
• AuditLog (real signing key generation, real append-only file).
• metrics_exporter_prometheus::PrometheusBuilder::install_recorder() (process-global, single-install).
• SpendTracker (real spend journal).

Mocking these would require introducing trait objects on the hot path, which is a non-trivial refactor that is out of scope for a tests-only PR. The pragmatic approach taken here:

• Test the isolated, testable surface area of the dispatch module without pretending we have the full pipeline under test.
• Document the gap explicitly so the next contributor knows where the seams are.

Coverage on src/server/dispatch/mod.rs improves from 0 -> 21 tests touching DLP scan, cache round-trip, audit entry construction, and result variants. End-to-end dispatch pipeline coverage is best added as integration tests in tests/integration/, where standing up a real AppState against a tempfile::TempDir is already common.

Note on base branch

Original task brief specified fix/preset-mod-include-str as the base, but that branch has since been merged into main (PR #304), so this PR targets main directly.

Test plan

• cargo nextest run --no-fail-fast — 1289/1289 pass (was 1268; +21 new).
• cargo nextest run --no-fail-fast --lib server::dispatch::tests — 21/21 pass.
• cargo fmt --check — clean.
• cargo clippy --tests --lib --no-deps — clean.
• No tokio::time::sleep in test bodies; deterministic by construction.