Fix apply: warnings must not mark a clean apply as failed

[lileding]

Problem

compose reports E_COMPOSE_APPLY_FAILED: <origin>: 0 op(s) failed [no per-op detail] for every dops-mode port, even though all ops apply cleanly (applied_ops == total_ops, errors == 0) and the generated output is correct.

Root cause

apply_plan (scripts/generator/dportsv3/engine/apply.py) computed:

ok = not diagnostics and all(row.status != "failed" for row in op_results)

not diagnostics requires the diagnostics list to be completely empty, but warning-severity diagnostics live in that same list. On a host without bmake, the default local oracle profile emits a single warning W_APPLY_ORACLE_SKIPPED ("bmake not found in PATH"). That lone warning flips ok to False, so compose reports the apply as failed — with zero failed ops and zero errors, hence the self-contradictory "0 op(s) failed" message.

This is not specific to any one port: it affects every dops-mode port on any host where the oracle emits a warning (e.g. bmake absent). It went unnoticed because the apply test-suite runs with oracle_profile="off", which never populates diagnostics.

Fix

Only error-severity diagnostics (or a genuinely failed op) mark the apply as failed:

has_error = any(d.severity == "error" for d in diagnostics)
ok = not has_error and all(row.status != "failed" for row in op_results)

Warnings are still recorded and surfaced; they just no longer masquerade as failures. Real oracle failures (E_APPLY_ORACLE_FAILED) and CI-profile unavailability (E_APPLY_ORACLE_UNAVAILABLE) remain error-severity and still yield ok=False.

Test

Adds test_apply_plan_local_unavailable_oracle_warning_keeps_ok: a clean apply whose only diagnostic is a warning-severity oracle skip must keep ok=True, with the warning still present. Verified red-before / green-after.

🤖 Generated with Claude Code