Error breakdown: error call paths keyed by RPC status (Python + Go) - #71

Error breakdown: error call paths keyed by RPC status (Python + Go)

Adds an error breakdown to both implementations: which call paths end in an error, keyed by RPC protocol and status code, with counts and exemplars.

crisp-trace --file t.json -s svc -a op --lightMode --errorBreakdown origins     # or propToRoot
crisp-go    --file t.json -s svc -a op --lightMode --errorBreakdown origins --errorBreakdownRoot trace

Writes error-breakdown.json next to the trace. The full spec is in the new CONFORMANCE.md section "Error breakdown".

Semantics

• Modes. origins reports every erroring span with no erroring child (it traverses successful spans too). propToRoot only reports errors that reach the root: a successful RPC span stops the walk. This is the existing Graph.computePropToRootGraph walk.
• Roots. trace (the default) starts at the trace's own root span. If no span lacks a parent, a virtual [INCOMPLETE_TRACE]virtual_root node parents the orphaned spans. analysis starts at the root chosen for -s/-a.
• Detection. A span errors if parseForErrorReturn flags it or its status code signals failure (HTTP >= 400, otherwise non-zero). Protocol and status come from standard tags (http.response.status_code, http.status_code, grpc.status_code, rpc.grpc.status_code, rpc.yarpc.status_code + rpc.transport, ...). A gRPC status name such as grpc.status = "UNAVAILABLE" sets only the protocol.
• Deployment hints. Hints specific to one deployment are new span_utils lists, all empty by default: HTTP_COMPONENTS, TCHANNEL_MARKER_TAGS, TCHANNEL_STATUS_TAGS, YARPC_STATUS_TAGS, IGNORED_ROOT_OPS.
• Tree used. The walk runs during Graph construction, before timeline sanitization drops out-of-range (clock-skewed) spans. Critical-path outputs are unchanged, and every existing conformance.* golden is byte-identical.

Fix

process() passed {} to getMetrics, so the --errorAnalysis flame graph of errors propagated to the root was always empty. It now passes computePropToRootGraph(), which also gains tests. --errorBreakdown outside light mode now logs a warning and is ignored, instead of being computed and never written.

Tests

• Goldens. Every conformance fixture gets error-breakdown-{origins,propToRoot}.json. Seven new test_cases/error_breakdown/ fixtures (status codes, orphans, multiple roots, a sanitized child, propagation cases) cover all four mode/root combinations. scripts/generate_goldens.py produces both sets and checks determinism.
• Parity. Python and Go both check the same goldens byte-for-byte. difftest -strict also compares error-breakdown.json when both templates pass the flag: 40/40 in both modes.
• Unit tests. Status-code extraction edge cases, the deployment lists, root selection, exemplar merging and counting, and option validation.
• Housekeeping. difftest -mode golden now checks only the conformance fixtures (as generate_goldens.py does). Gazelle added the missing light_test.go to go/crisp:crisp_test.

Validation: scripts/ci-local.sh (837 passed, test.sh clean), bazel test //... (37/37), go vet ./..., and difftest in golden mode (33/33) and strict corpus mode in both breakdown modes (40/40).

About 7k of the added lines are golden JSON.