fix(indexer): cap large document assets and skip vector data by default

[zzet]

Summary

Fixes the content-repo OOM in #120. A RAG corpus (the reporter's repo: ~2,045 admitted files / ~4.42 GiB, ~85% pptx/xlsx/pdf, ~99% non-source documents + vector data) could exhaust memory during indexing because every document and data artifact was admitted into the read+extract pipeline with no per-class cap and no opt-out — so a few hundred large non-source files pulled gigabytes into memory.

Prior fixes bounded the symptom (#132 large-read gate, #137 bytes-in-flight semaphore + shadow-bytes guard + content/code split) but not the cause: there was no corpus admission control. This adds one.

What changed

A walk-time corpus-admission gate drops non-source artifacts before they are read and extracted:

• Non-code extractors declare an AssetClass (document / data / image) via a new parser.AssetExtractor marker interface. The indexer builds a language -> class map from the registry once per walk (inert for all-code repos).
• Documents (pdf/pptx/xlsx/text) over a per-file cap (default 10 MiB) are skipped with skip_reason: large_document.
• Binary/vector data (parquet/npy/npz/lance/arrow/feather) are skipped entirely by default with skip_reason: vector_data.

Skipped files become synthetic KindFile nodes (skipped_due_to_content) so they stay listable and roll up in index_health, mirroring the existing size/timeout skip telemetry. The gate runs on both the cold IndexCtx walk and the incremental indexFile (watcher) path, so the live path can't re-admit what the cold path skipped.

Markdown prose and code are never gated.

Configuration

All behaviour is configurable under index.content:

index:
  content:
    max_document_bytes: 10485760   # >0 cap | 0 default | <0 no cap
    index_data: false              # opt in to admit data assets
    max_data_bytes: 67108864       # cap applied only when index_data is on

Caps are tri-state (>0 cap / 0 built-in default / <0 no cap), so a zero-valued config still yields the correct defaults regardless of how it was constructed. Users who want the prior behaviour set max_document_bytes: -1 and index_data: true.

Behaviour change

This changes defaults for existing users: documents over 10 MiB and all vector/data artifacts are no longer indexed unless opted in. This is the intended outcome per the issue discussion — the multimodal feature stays intact for normal-size documents.

Validation

• go build ./... (CGO) — success
• go test -race ./internal/config/... ./internal/parser/... — 1942 passed
• go test -race ./internal/indexer/ — 668 passed (incl. new content_admission_test.go)
• go vet + golangci-lint on changed packages — clean
• cmd/gortex wire-contract golden — unaffected (no graph node/edge field change)
• Confirmed no test fixtures regress (data-asset files only exist under eval/.venv/)

Notes

The in-memory gortex init path (graph.New, nil content sink) still retains full section text; the admission gate shrinks how much it accumulates, but unconditional content-node leaning was intentionally skipped to avoid regressing in-memory content search on the --backend memory case. Left as a possible follow-up.

Closes #120