14 cross-service protocol linkers (PR 2/5)

[Shidfar]

Summary

Adds the 14 protocol linkers that populate the empty dispatch table introduced by #376.

Stacked on #376 — please review that first. Until #376 merges, this PR's diff includes both #376 and PR 2 commits; it'll narrow to PR-2-only after #376 lands.

Commits

1. feat: add GraphQL and gRPC protocol linkers
2. feat: add Kafka, SQS, SNS, and EventBridge protocol linkers
3. feat: add Pub/Sub, RabbitMQ, MQTT, NATS, and Redis Pub/Sub linkers
4. feat: add WebSocket, SSE, and tRPC protocol linkers
5. build: wire 14 protocol linkers into pipeline — restores the full LINKERS table in pass_servicelinks.c, adds source files to PIPELINE_SRCS, registers the 14 suite_servicelink_* tests, and fixes a pre-existing latent UB bug in cbm_pipeline_run (see below)

Pre-existing UB bug fix in cbm_pipeline_run

The plumbing commit in #376 added cbm_sl_endpoint_list_free((cbm_sl_endpoint_list_t *)ctx.endpoints); to the cleanup: block. But ctx is declared after the early-cancel check (if (rc != 0 || check_cancel(p)) { goto cleanup; }), so when cancellation fires before any pipeline work, the cleanup block reads uninitialized stack memory through ctx.endpoints.

I caught this with ASan in test_integ_pipeline_cancel: the uninitialized slot happened to contain a stale sqlite-freed address from the previous test_integ_mcp_list_projects (sqlite's pcache1EnforceMaxPage freed memory during cbm_store_close, and that address landed in ctx.endpoints's slot on the cancel test's stack frame), surfacing as a heap-use-after-free in cbm_sl_endpoint_list_free. It's latent UB — the full oss/clean-features stack passed because subsequent commits' stack-layout shifts happened to put NULL/safe values there.

Fix: declare cbm_sl_endpoint_list_t *endpoints = NULL; at the top of cbm_pipeline_run alongside path_aliases, allocate the list after the cancel check, assign to ctx.endpoints, and free endpoints (not ctx.endpoints) in the cleanup block. Safe regardless of which goto fires, since endpoints is declared before all gotos and initialized to NULL (and cbm_sl_endpoint_list_free(NULL) is a no-op).

Test plan

• ./scripts/test.sh passes (3796/3796, ASan + UBSan)
• All 14 suite_servicelink_* test suites green
• ASan no longer reports use-after-free in test_integ_pipeline_cancel

Upstream overlap audit (re-checked against upstream/main @ 6226972)

Since this PR was opened the audit has been re-run on current upstream. Findings:

• Already covered upstream (12 of 14 protocols):
  • internal/cbm/service_patterns.c (lines 152-248) — library detection tables for kafka, sqs, sns, eventbridge, pubsub, rabbitmq/amqp, nats, redis, mqtt, grpc, graphql, trpc
  • src/pipeline/pass_calls.c:253-257 — emits HTTP_CALLS and ASYNC_CALLS edges with broker metadata
  • src/pipeline/pass_parallel.c:1221-1327 — emits GRPC_CALLS, GRAPHQL_CALLS, TRPC_CALLS edges
  • src/pipeline/pass_cross_repo.c:match_typed_routes (line 492) — cross-repo typed-route matching for gRPC/GraphQL/tRPC
  • src/pipeline/pass_cross_repo.c:match_async_routes (line 330) — cross-repo matching for messaging protocols
• Net-new in this PR:
  • WebSocket and Server-Sent Events (SSE) linker coverage — no upstream equivalent
  • Regex-source-scan enrichment for the other 12 protocols that produces topic/exchange/QoS metadata not derivable from upstream's library-name + call-resolution path
• Recommended path: trim the 12 overlapping protocol files; keep WS + SSE only. Alternatively, migrate the 12 to enrichment hooks layered on top of upstream's existing *_CALLS edges rather than emitting them in parallel.

Marking remains draft until reviewed against this audit. PR #380 establishes the architectural reconciliation (cedes 4 protocols to upstream); the consolidated shape of this PR depends on how that lands.