feat(standing_rules): inject_count + last_injected — named for what it measures

[AVADSA25]

Requested as fired_count. Shipping it as inject_count, because "fired" is not measurable here — and the requester pre-authorised that fallback if so.

Where a rule is genuinely "applied"

The complete lifecycle:

prompt_block() → string → sys_prompt += block → sent to the LLM

That is the whole mechanism. There is no downstream point where a rule fires. Nothing in the pipeline observes whether the model consulted it, obeyed it, or ignored it.

Measuring real influence would require:

• ablation — run the turn with and without the rule and diff; doubles cost, and non-determinism makes a single diff weak evidence
• asking the model which rules it used — precisely the unverifiable self-report codec_claim_check exists to refuse

A field named fired_count would be a fabricated metric wearing an authoritative name, in the module whose entire job is refusing that.

What the number is actually good for

Value	What it proves
inject_count == 0	dead weight — the rule never even reached the model. Safe to delete.
high inject_count	nothing about usefulness — only that you pay for it on every message

The listing states this in the user-facing text rather than implying the model used anything, and calls out rules that have never reached the model as safe to delete.

Design details

• record=True is passed only by the live chat path — listing, preview and tests default to False, so the metric cannot measure itself
• Migration is additive: existing rules keep id/text/added and gain inject_count: 0 / last_injected: null. The store is never reset.
• Telemetry write failures are caught — a counter must never break a reply

Verification

Migration on a real pre-existing store:

before: {'id': 'old1', 'text': 'Always answer in French.', 'added': '2026-07-01T10:00:00+00:00'}
after : {... 'added': '2026-07-01T10:00:00+00:00', 'inject_count': 0, 'last_injected': None}

Three injections → counts [1,1] → [2,2] → [3,3], persisted to disk with timestamps.

20 standing-rules tests, including "preview does not count itself", "count survives a reload", and one asserting the listing text never uses the word "fired".