fix(cmdfilter): bound selector-miss ledger keys by size, and drop non-text blocks

[OsherElhadad]

Closes #48. Found by the integration review, replaying real captured traffic through merged main.

Problem

maxMissKeys = 200 bounds how many keys the ledger holds, not how big they are — and selectorKey runs on whatever the tool returned. Replaying capture-tb.jsonl filled the top 25 slots with base64 image payloads:

[{"type":"image","source":{"type":"base64","data":"iVBOR…

Two consequences:

1. The ledger stops doing its job. It exists to answer "which filter is worth writing next" (after rtk's parse_failures table). On multimodal traffic that answer is 200 near-identical image blobs, crowding out the actionable misses — the real ones observed were Exit code 1 (211×) and source listings.
2. 200 multi-KB keys sit in the aggregator under its lock and ship in every /stats scrape.

Fix

Bound the key at 120 bytes. A selector is a shape, so a short prefix identifies it; the tail only makes near-identical misses occupy separate slots. Well above any real command banner, well below a payload. Truncation cuts on a rune boundary so a key stays valid UTF-8 in the JSON payload.

Drop non-text blocks entirely rather than truncating them. An image carries no output shape a filter could ever match, so recording it is noise by construction — not a key that merely happens to be too long. A truncated blob would still be a useless entry occupying a slot.

Test

TestMissLedgerKeysAreBoundedAndTextOnly asserts four things, and the fourth is the one that matters most:

• the key is bounded, and keeps its identifying prefix (docker build … survives the cut);
• a truncated key is valid UTF-8 (tested with multi-byte runes);
• non-text blocks are dropped, in both compact and spaced JSON spellings;
• real command banners survive intact — Reading package lists..., > Task :app:compileDebugKotlin, make[1]: Entering directory '/src'. The bound must not cost the signal it exists to rank, and a test that only checked the truncation would happily pass a version that mangled every real selector.

Verified non-vacuous — removing the non-text guard reproduces the reported symptom exactly:

--- FAIL: TestMissLedgerKeysAreBoundedAndTextOnly
    non-text block recorded as a miss shape: "[{\"type\":\"image\",\"source\":{\"type\":\"base64\",\"data\":\"iVBORw0KGgo…

Gates

go build -tags cg_skeleton ./... · go test -tags cg_skeleton ./... · gofmt -l — all clean.

Remaining integration findings

#45 (the volatile-tail split is a silent no-op on Bedrock Converse — it reports Changed: true and achieves nothing) and #47 (pin-budget exhaustion reopening the TailOnly fail-open, on the /compact route only) are still open.