feat(indexer): split benign code-pipeline file skips out of errors_total

Splits benign per-file skips out of gkg_indexer_code_errors_total into a new gkg_indexer_code_files_skipped_total{reason} counter so policy outcomes stop dominating the error rate.

In orbit-prd over 24h, errors_total{stage="parse"} showed 1111 events while real failures (repository_fetch) showed only 7. The parse bucket is filled by the JS pipeline emitting record_error after warning js: skipped file for files larger than 2 MiB or with lines longer than 5000 bytes. The per-file watchdog (Internal { context: "sentinel_timeout" }) also flowed through the same counter. None of these are failures.

What changed

• New metric gkg.indexer.code.files.skipped (Prom: gkg_indexer_code_files_skipped_total), label reason, declared in crates/gkg-observability/src/indexer/code.rs.
• Pipeline context grows record_file_skipped(path, reason) next to record_error. Skips drain into PipelineResult.files_skipped (crates/code-graph/src/v2/pipeline.rs).
• JS pipeline classifies per-file analyzer messages: "refusing oversize file" -> oversize, "line too long" -> line_too_long, anything else stays on record_error (crates/code-graph/src/v2/langs/custom/js/pipeline.rs#L62-75).
• Rust pipeline uses the same classifier symmetrically (crates/code-graph/src/v2/langs/custom/rust/mod.rs#L146-157).
• Per-file watchdog reroutes from Internal { sentinel_timeout } to files.skipped{reason="timeout_sentinel"} (crates/code-graph/src/v2/pipeline.rs#L982-994).
• Indexer dispatch increments the new counter from the drained list (crates/indexer/src/modules/code/indexing_pipeline.rs#L294-301).
• CodeGraphError::stage() is unchanged; only the recording side moves. Genuine grammar failures still surface as errors_total{stage="parse"}.

Tests and checks

• New unit tests in js/pipeline.rs write an oversize file and a line-too-long file and assert files_skipped increments while graph_errors stays empty.
• Pipeline-level tests cover classify_skip_message and the record_file_skipped collection path used by the watchdog.
• mise test:fast passes (1900/1900).
• mise lint:code clean.
• mise run metrics:catalog and mise run dashboards regenerated; pre-commit metrics-catalog-check and dashboards-check both green. The auto-generated reference panel for the new metric appears in orbit-gkg-indexer.dashboard.json and orbit-all-metrics.dashboard.json; no story-shaped panels were added.
• docs/design-documents/observability.md lists the new metric and notes the rationale.

Closes #529 (closed) Refs &20992