fix: explain why an element reference stopped resolving

[yashranaway]

Closes #16.

The problem

snapshot() reissues element references on every call, so a reference from an earlier inspection silently stops resolving. The common case is nasty: run --context full (200+ controls), then --context summary (8 controls), and every reference from the first call is now dead. The agent received:

ELEMENT_NOT_FOUND:@e137

Nothing there distinguishes "this expired, re-inspect" from "this never existed" — which is exactly the difference between recovering and retrying a dead reference forever.

The fix

Failures name the cause:

Cause	Meaning
expired	issued earlier, superseded by a later inspection — inspect again
unknown	never issued — do not guess or invent references
detached	the element or region has left the page

Both ELEMENT_NOT_FOUND and REGION_NOT_FOUND carry it. Host error-code mapping is unchanged (it matches on the prefix), so this is additive on the wire.

The asymmetry is deliberate

I kept element and region lifetimes different, and documented why rather than "fixing" it:

• @eN is latest-inspection-only. That is already what the skill teaches, and making references durable would mean holding every control on every page.
• @rN stays resolvable across inspections, because the documented outline-then-scope workflow depends on it — take @r4 from an outline, then scope repeated text/actions calls to it. An existing test covers that workflow; changing it would have broken a shipped contract.

What was wrong with regions is that the map was never cleared, holding strong references to detached nodes for the lifetime of the page. It is now bounded at 256, oldest-first — a leak fix, not a semantic change.

Verification

Ran the runtime suite locally (this one I can run — it is Node-only): stale reference reports expired, unknown reference reports unknown, a fresh reference still resolves, and the cross-inspection region workflow still works. The pruning numbers are unchanged (48,428 → 895 bytes), so the headline metric did not regress.

P1 gains a Reference lifetime section so this is documented behaviour rather than folklore.