Token-saving observability

fak serve and fak manage expose a bounded token_savings receipt on the existing read-scoped GET /debug/vars endpoint. Loopback callers need no bearer; remote callers need the read bearer. The receipt contains aggregate counts and bounded reason enums only: no prompt text, tool arguments, paths, traces, or tool results.

(Invoke-RestMethod http://127.0.0.1:8080/debug/vars).token_savings |
  ConvertTo-Json -Depth 6

Each default-on lever reports:

• configured: whether the resolved gateway posture armed it;
• state: active, ready (armed but not yet exercised), bypassed, or off;
• reason and bail_reasons: why it fired or safely returned identity;
• fired, units, and measured saved_bytes or saved_tokens where that transform has a truthful effect measure;
• rollback: the exact flag or environment switch that restores the baseline.

The native MCP filter additionally emits the same receipt in every tools/list response at _meta["fak/tool_filter"]. Its active/control proof is executable as go test ./internal/gateway -run TestNativeMCPFilterABProof: the active arm must reduce exact descriptor bytes, the FAK_ABLATE_MCP_TOOL_FILTER=1 arm must restore the full list, held intent queries must retain recall, and an --expose-hidden tool must remain both undiscoverable and uncallable.

Interpret ready/not_observed as not yet proven to have saved anything in this process, not as a win. A bypassed state is an intentional safe bailout: fak kept the original request/tool surface and paid the token cost rather than risking a missing capability or an ambiguous rewrite.

Audit effectiveness coverage

Use the defaults scorecard’s effectiveness view to see how every default saver is measured and how to restore its control arm:

fak token-defaults-scorecard --effectiveness
fak token-defaults-scorecard --effectiveness --json

The table is source-derived from the same nine-saver roster as the defaults scorecard. It names the correct witness class for each mechanism: same-trace fak ablate, exact-byte A/B, live receipt, or suite benchmark. A newly added default saver without an effect metric, witness, and control becomes audit debt instead of inheriting a “proven” claim from merely being configured or regression-locked.