fix: declare unique_constraint on audit_log pkey to prevent Ecto.ConstraintError in *_and_log (OPS-4594)

[palantir-valiot]

Summary

Add a unique_constraint(:id, name: <table>_pkey) declaration inside changelog_changeset/1 so the bare repo.insert/1 performed by log_changes/5 (and log_changes_alone) for every *_and_log / log path turns a duplicate pkey violation into a changeset error instead of letting Ecto raise Ecto.ConstraintError.

The error path already existed and was intentionally swallowed (logs the reason and returns {:ok, reason}), so callers continue to see success for the primary operation while the audit write is best-effort.

Files changed:

• lib/ecto_trail/ecto_trail.ex:574 (core fix + two small module attrs for the constraint name, derived from the same table_name config the schema already uses)
• test/unit/ecto_trail_test.exs (TDD test that forces a deterministic audit_log pkey collision on the update_and_log path and asserts the main update still succeeds)
• mix.exs (1.0.3 → 1.0.4)
• CHANGELOG.md (concise entry)

Why

Production (eliot-lamosa-gto-prod) hit:

(Ecto.ConstraintError) constraint error when attempting to insert struct:
    * "audit_logs_pkey" (unique_constraint)
...
    (ecto_trail 1.0.3) lib/ecto_trail/ecto_trail.ex:435: EctoTrail.log_changes/5
    (ecto_trail 1.0.3) lib/ecto_trail/ecto_trail.ex:315: anonymous fn/4 in EctoTrail.update_and_log/4

Root cause: changelog_changeset/1 only did Changeset.cast/3; no unique_constraint/3 was ever declared for the pkey of the (configurable) audit table. Under concurrent load the sequence can emit a value that another writer has just used, producing the exact pkey collision visible in the stacktrace.

See Linear OPS-4594 and the triage analysis (NOTIFY+FIX, code_bug).

Assumptions

• The audit table always has a serial/bigserial id pkey whose Postgres constraint name is <table_name>_pkey (the default audit_log produces audit_log_pkey; the observed prod name was audit_logs_pkey because some tenants use a plural table name via config). The constraint name is computed the same way the schema already derives the table name.
• The existing "swallow log errors" behavior in log_changes/5 and log_changes_alone is intentional and sufficient; we only need to stop the exception from escaping.

Test plan

• mix format (clean)
• mix compile (clean; only pre-existing redundant-clause warning on an unrelated helper)
• Added a focused regression test under describe "update_and_log/3 pkey constraint resilience" that:
  • Seeds a colliding row with a known id
  • Uses setval so the next nextval() inside the library-produced insert will collide on that pkey
  • Calls update_and_log
  • Asserts {:ok, updated_resource} (no ConstraintError) and that the main resource row was mutated
• mix test — the suite has always required a running Postgres (see test/test_helper.exs:82). In this k8s per-task pod no Postgres listener is available, so the harness cannot reach the "migrate + sandbox" steps. The compile-time and formatting gates pass; the new test was executed mentally against the exact failure mode from the stacktrace and will be exercised by CI on PR (or by any developer with docker compose up / the project's normal test DB). No other verification scripts are declared in mix.exs, AGENTS.md (absent), or package.json (none).
• Self-review of git diff origin/main..HEAD: only the four files above; no debug prints, no unrelated refactors, no new deps, no scope creep.
• PR body written from scratch (no .github/pull_request_template.md present; skeleton not used verbatim).

Closes OPS-4594

[linear-code]

OPS-4594

[acrogenesis]

Closing as a duplicate of #24 — same bug: Ecto.ConstraintError on audit_logs_pkey in EctoTrail.log_changes/5. These were generated from a backlog of duplicate Linear issues created by a log-agent dedup gap (now fixed in palantir 38438d6; no new duplicates are being filed). Consolidating on #24.