chore(metrics): freeze the proc-v1 process digests - #991

Holding this merge deliberately — it is the freeze cutover, and the pre-freeze queue now assigned to the OT2026 milestone (#897, #903, #904, #963, #967, plus #864's remaining verification) lands first. Two notes for whoever merges it:

1. Re-paste the digests before merging if any of the queued work moves a prompt or config that feeds compute_process_digest — metrics: decide how segment_base_rate reconciles a sal-v2 band against a lagging statpack #897 (evaluate.md's segment_base_rate text), evaluate: score predicted judicial reasoning (two-tier) #903/analytics: surface predicted-reasoning scores, and validate the judge against Tier 1 #904 (a claim set touches both prompts), and salience: sal-v1 conditions on signals a petition does not have on arrival — rework for prospective use #963 (predict.md names sal-v1 verbatim) all can. Re-run fedcourts process-digest --all at the final content and update the FROZEN_PROCESS_DIGESTS paste; the second commit's milestones.md sha reference must then also be redone (which is fine — merge commit, not squash, per the description).
2. The "closes … alongside metrics: prove the leaderboard consume path on production evaluations #864" line is premature. The frozen-scope leaderboard is currently empty (entries: [], evaluations_total: 0), the evaluate channel is paused with 3 queued events held by the handoff variable, and mean_brier_skill_score has never produced a number. metrics: prove the leaderboard consume path on production evaluations #864's own rescope comment names what remains: resume the handoff, let the three held cells grade, hand-check each segment_base_rate against the pre-computed sheet. metrics: prove the leaderboard consume path on production evaluations #864 stays open past this promotion until that check is done — it is itself a pre-freeze validation of the metric the freeze makes claimable.

The pasted digests are stale — do not merge this as it stands

All six values in FROZEN_PROCESS_DIGESTS predate the pre-freeze batch. Two merged changes moved every actor's digest: #1013 (the cert-v1 claim contract in predict.md / evaluate.md) and #1024 (the redaction change), plus #1022 and #1027 (the stage-conditional prompt sections). The digest is sha256(prompt_bytes ␀ canonical{engine, resolved model, prompt_path, retrieval, mcp_servers}), so a prompt edit moves it by construction.

The PR's own frozen-state test is the backstop here — it requires the set to resolve exactly against the current registry, so merging the stale paste fails the suite rather than silently freezing the wrong process. Worth stating anyway, because the failure would read as a broken test rather than a stale paste.

fedcourts process-digest --all on the current staging tip (63bef3b3) produces:

proc-v1  predictor  claude-baseline  sha256:30c4b708a2a28e608aa906858758aa2cfe972c3b36a0dd15952f6c1db4f9775d
proc-v1  predictor  codex-baseline   sha256:b80f2290bfb29c12801e84d9b4a60dc992426a7a9eb91b85dd0a7c1835968ed8
proc-v1  predictor  gemini-baseline  sha256:928552054ef6deb61dbf1f7f5648a7eb66f0bfd4b52d646aa9746f69a29fecf2
proc-v1  evaluator  claude-judge     sha256:7ed1a24d1b4cc7fea57f229a29ecc700e980eccd7e2d2b7b071412a343581c4a
proc-v1  evaluator  codex-judge      sha256:2fe642cf0a84065c129241bdd053efdb4e243b524b78b8cfe747e3953ef15047
proc-v1  evaluator  gemini-judge     sha256:56c9cfd91024512ee24af3fe34f312dc0e279b602e28b9e0a0e3b8f443512b77

These are a preview, not the paste. sal-v2 (#963) is still to land and the promotion has not happened, so the sequence stays: land sal-v2 → promote (#1019, merge commit) → run fedcourts process-digest --all on the promoted tree → paste that output here → merge-commit → bless.

The "Pre-freeze evidence" section above is also stale: the all-scenario green it cites was at staging 4b93895e / 802ea80e, many merges back. The evaluate-side engine smoke it lists as outstanding is still outstanding.