Design read-only identity verification without breaking wait polling

[thymikee]

Problem

ADR 0012 target-binding evidence and replay-time refusal do not yet cover every selector-targeted read-only or synchronization step.

A destination guard in ADR 0016 is a polling wait <selector>. It currently proves selector existence, not that the resolved landmark has the same identity observed during authoring. A reshuffled screen containing the same label elsewhere can false-pass.

The existing ADR 0012 mechanism cannot simply be attached to wait: it verifies target-v1 before dispatch. A wait commonly begins while its landmark is absent, so pre-action verification would reject the expected starting condition instead of polling for it.

get already passes recording-time node/tree evidence on its recording-aware snapshot path. This work should preserve that behavior while defining compatible semantics for waits and the remaining eligible read-only steps.

Desired outcome

Design and implement recorded identity verification for eligible read-only steps without changing their execution semantics. For wait, selector polling must happen first; any recorded-landmark identity comparison happens only after a candidate resolves and before the wait reports success, or uses a distinct guard-specific mechanism with the same fail-closed outcome.

Acceptance criteria

• wait <selector> continues polling while no candidate exists; identity verification never performs an eager pre-action refusal merely because the landmark has not appeared yet.
• When an annotated wait resolves a candidate, an identity mismatch/unverifiable result fails closed through the bounded REPLAY_DIVERGENCE contract before the wait reports success.
• The design explicitly chooses whether wait evidence can reuse target-v1 with wait-specific dispatch or requires a distinct bounded annotation; it must not accidentally enter the existing generic pre-action path.
• Duration waits and wait stable remain unannotated because they have no element target.
• wait @ref is not treated as a portable destination guard; ADR 0016 publication accepts only a selector or selector chain.
• Eligible is and read-only find variants are explicitly classified as covered or intentionally deferred, with tests for each decision.
• Existing get identity behavior remains intact and shares ADR 0012's verification machinery where its immediate-read semantics permit pre-action verification.
• ADR 0016 gains a reshuffled-screen false-pass regression when wait identity verification ships.

Relationship

• Follow-up to ADR 0012's bounded target identity model.
• Strengthens ADR 0016 / Publish an open-to-destination script without closing the session #1346 destination guards from selector existence to recorded landmark identity without breaking polling.
• Corresponds to the replay-study roadmap item “extend identity annotations to read-only steps.”

[thymikee]

Triage update: this is a valid follow-up, but it is not ready-for-agent yet. Before implementation, specify that wait identity verification happens after the wait resolves its landmark (not as the existing pre-dispatch verification that would prevent waiting for a newly appearing target), define how a recorded wait @ref is converted to a portable selector or rejected, and explicitly classify the eligible is/read-only find variants rather than leaving that decision to the implementer. Keep it sequenced behind acceptance of ADR #1347; #1349 strengthens identity but does not replace ADR 0016’s portable-guard contract.

[thymikee]

Triage update after the rewrite: the issue now correctly preserves wait polling, places identity verification after candidate resolution, excludes wait @ref as a portable guard, and requires explicit read-only command classification. It is sufficiently bounded for implementation, but remains intentionally unlabeled until ADR #1347 is accepted/merged; add ready-for-agent once that dependency lands.