refactor(metrics): re-home metrics onto aragora.evaluation with shims (VAL-P4A-012)

[scarmani]

What

Re-home the aragora/metrics/ originals onto aragora/evaluation/ and leave
aragora/metrics/*.py as DeprecationWarning re-export shims so every legacy
import path keeps working for one release (VAL-P4A-012).

Moved (pure git mv, content unchanged) to aragora/evaluation/:

• viah.py, viah_signals.py, viah_status.py
• manifold_brier.py, manifold_brier_bridge.py
• capability_checkpoint.py

aragora/metrics/ now contains only two-sided re-export shims:

• each submodule re-exports its __all__ from the matching
  aragora.evaluation.* home and emits a DeprecationWarning whose message
  names the old aragora.metrics.* path
• the package __init__ emits an aragora.metrics is deprecated warning and
  keeps its lazy PEP 562 __getattr__ (now resolving against
  aragora.evaluation)

Intra-package imports inside the two moved bridge/status modules were repointed
to their new aragora.evaluation.* homes. The only external importer,
aragora/cli/commands/agt_metrics.py, intentionally stays on the shim.

Behavior

• All documented old import paths (from aragora.metrics import ...,
  from aragora.metrics.viah import ..., etc.) keep working and now emit a
  DeprecationWarning.
• Re-exported objects are identical to the aragora.evaluation.* originals
  (aragora.metrics.viah.compute_viah is aragora.evaluation.viah.compute_viah).

Validation

VAL-P4A-012 two-sided shim check (every tracked aragora/metrics/*.py):
code-free shim (only __getattr__/__dir__ allowed), re-export restricted to
aragora.evaluation/aragora.metrics, old path import works, new home
importable, package DeprecationWarning contains metrics:

########## SUMMARY FAIL=0 ##########

Local gates: make lint PASS; ruff format --check aragora/ tests/ scripts/
PASS; CI changed-file typecheck (mypy --ignore-missing-imports --follow-imports=skip over the 13 changed files) Success: no issues found;
tests/metrics/ + tests/cli/test_agt_metrics_status.py 234 passed (one
pre-existing failure test_invalid_weeks_raises, confirmed identical on clean
origin/main, unrelated to this move); make test-smoke and the smoke tier
(138 passed) PASS.

Notes / risks

• docs/METRICS.md and aragora/module_tiers.yaml are intentionally NOT
  regenerated here: both are path-frozen by open PRs chore(deps): update audit-vulnerable Python locks #8460/fix(ci): clear base security and smoke blockers #8461/feat(tier4): settle_tier4_pr.py — one-command --settle-apply + receipt-backed preapproval gate #8382 and are
  handled by the phase merge-train's single consolidated regen.
• Touching aragora/metrics/** makes this a Tier-3 PR (path-based). Prepared
  for operator risk settlement.

[github-actions]

Aragora Code Review

Advisory-only review. No issues found.

[scarmani]

Claude independent model review

Reviewer: claude (anthropic) — independent adversarial model review via the Aragora Claude reviewer, grounded on the exact PR head.
Head: 6d69a0d (6d69a0d), committed 2026-06-29T18:12:56Z.
PR: #8696.
Model family: claude

Verdict: PASS

Symbol parity and mechanics check out: every shim re-exports the full public surface of its new aragora.evaluation.* home (viah 18 names, manifold_brier 6, manifold_brier_bridge 5, viah_status 1, capability_checkpoint 8 incl. build_default_registry), all evaluation modules are self-contained (no back-references to aragora.metrics), all 13 files parse, and the moved code is logic-identical to the originals. The package __init__ __getattr__ redirect targets were correctly repointed to aragora.evaluation.* with no change to its __all__. No filterwarnings=error in pytest config, so the added DeprecationWarnings won't break the suite.

• [P3] aragora/cli/commands/agt_metrics.py:15,71 — first-party live consumer still imports the now-deprecated shims (from aragora.metrics.viah import ... and aragora.metrics.viah_status). Functionally fine (shims re-export correctly), but the project ends up deprecation-warning against itself on its own agt-metrics CLI path; the shims exist for legacy/external callers, so this internal caller should be repointed to aragora.evaluation.viah / aragora.evaluation.viah_status in this PR.

dogfood: yes

[scarmani]

OpenAI independent model review

Reviewer: openai (openai) — independent adversarial model review via Codex CLI OpenAI harness, grounded on the exact PR head.
Head: 1e0c4c1 (1e0c4c1), committed 2026-06-29T20:01:35Z.
PR: #8696.
Model family: openai

Verdict: PASS

No findings.

dogfood: yes

[scarmani]

Claude independent model review

Reviewer: claude (anthropic) — independent adversarial model review via the Aragora Claude reviewer, grounded on the exact PR head.
Head: 1e0c4c1 (1e0c4c1), committed 2026-06-29T20:01:35Z.
PR: #8696.
Model family: claude

Verdict: PASS

This is a clean re-homing refactor: the metric modules (viah*, manifold_brier*, capability_checkpoint) move verbatim to aragora.evaluation.*, and aragora.metrics.* becomes deprecation shims that re-export them. Verification results:

• aragora/evaluation/__init__.py lazy __getattr__ is correct — all 9 names in __all__ are defined in llm_judge.py; from __future__ import annotations is correctly placed right after the docstring.

• The new lazy init does what it claims: importing aragora.evaluation.viah does not pull in aragora.evaluation.llm_judge (confirmed at runtime → the new test_viah_import_does_not_load_llm_judge_stack passes).

• All shim re-exports are complete (viah.py exports ViahTrend/persist_viah_snapshot/read_viah_snapshots/etc.; capability_checkpoint exports build_default_registry; package-level __getattr__ resolves compute_viah/ManifoldBrierScorer). Confirmed by importing each.

• No remaining production importers of aragora.metrics outside the shims themselves, and no filterwarnings = error in pytest config — so the new module-level DeprecationWarning won't break CI. The pattern matches existing in-repo shims (type_protocols, redis_config).

• The removed assert report.window_start.endswith("Z") line in test_viah.py was unreachable/dead (after a pytest.raises block, referencing an out-of-scope report); removing it is correct.

• [P3] aragora/metrics/*.py (all shim files) — the legacy aragora.metrics.* import paths have no remaining test coverage in this PR (tests were migrated to aragora.evaluation), so a future typo in a shim re-export list would not be caught by CI before the "one release" back-compat window closes. I confirmed they all import/re-export correctly today; consider a tiny smoke test asserting the shim paths still resolve.

dogfood: yes