fix(gp): bless live response oracles for 9 unblessed scripts + gate future regrowth - #229

Stacked on #228 (its current head 7309939), because trunk's ephemeral-server-smoke is red. The unpinned seed inserts services.max_record_count, but the nightly-aot image schema has no such column, so the seed fails with psql exit 3. Once #228 merges, this diff shrinks to its two commits.

Problem (#202)

ephemeral-server-smoke graded no response oracle for 9 of the 25 supported honua-gp eval scripts:

• describe_segments[_fields]
• list_fields_segments[_filtered|_wildcard]
• insert_cursor_append_rows
• update_cursor_close_status
• update_cursor_delete_closed
• make_table_view

A supported script with no oracle counted as a silent "unblessed" pass. So a regression in Describe/ListFields or in the write cursors could pass this lane unnoticed. --require-supported-pass-rate 1.0 did not help, because unblessed scripts counted as passing.

Changes

Commit 1 (the earlier revision, rebased onto #228). Blessing the cursor scripts exposed a real bug. honua_gp.da read only .attributes, but the live SDK's QueryFeature exposes .properties/.id. As a result, every live SearchCursor/UpdateCursor field read, OID included, returned None. This commit:

• fixes that bug, with regression tests against QueryFeature;
• adds schema oracles for Describe/ListFields;
• adds apply-edits oracles for the cursor scripts.

Commit 2 (this revision). Commit 1 still left three gaps:

• update_cursor_delete_closed was blessed to delete_count: 0. An earlier script had already archived every CLOSED row, so it deleted nothing, and a broken deleteRow would have passed.
• The write-cursor oracles recorded only applyEdits counts, never the dataset contents afterwards.
• make_table_view still passed through an allow-list as "unblessed".

Commit 2 closes them:

• Insert. insert_cursor_append_rows reads back the object ids the server assigned and records the rows that persisted.

• Update and delete. update_cursor_close_status and update_cursor_delete_closed each insert their own fixture row and scope the edit to it by name. They then read back:

  • the value the update left;
  • that the deleted id is the inserted id;
  • that zero rows remain.

  Neither depends on script order or on rows left by an earlier run.

• Table view. make_table_view applies status = 'active' and records the rows the server returns through the view (GetCount + SearchCursor).

• Gate. The allow-list is removed. In live mode, a supported script without a response oracle now always fails. A new unit test, test_every_supported_script_has_a_response_oracle, fails the stub lane as soon as a supported golden lacks one.

• Generator. _generate_scripts.py now owns these four scripts' oracles. Before this change, regenerating the scripts silently dropped the hand-edited emit lines.

Per-script decision

All 9 scripts have an observable result, so the unblessed set is now empty.

Evidence

All live runs used the pinned candidate: ghcr.io/honua-io/honua-server@sha256:dd50cd81… with the seed at 7ba4226. The stack was started with the workflow's own compose override (migrate first, then seed, then the SpatialJoin join layer).

• Fresh stack, committed goldens, at head 3b46f68: honua-gp eval [live]: 52/52 passed (100%); supported 27/27 (100%).

• Bless run on a fresh stack: only the four scripts changed in this revision produced new goldens. The other 48 matched their committed goldens.

• Rerun against the already-mutated database: all four changed scripts pass again. The older seed-pinned count scripts drift there as expected, as docs/golden-eval.md already documents.

• Negative control: I reverted only the .properties fallback in _attrs_for_feature and ran just the four changed scripts live. The result was supported 0/4:

  • update_cursor_delete_closed got delete_count 0, rows_remaining 1. The old count-only oracle blessed exactly that as a pass.
  • make_table_view got [[null, null] × 5].
  • The two remaining cursor scripts also failed their read-back values.

  I restored the file afterwards and the worktree was clean.

• Unit tests: python -m pytest packages/honua-gp/tests -q gives 281 passed, 9 skipped.

• Other checks: stub eval 52/52; python -m honua_gp._cli matrix --check packages/honua-gp/docs/compatibility-matrix.md passes.

• Out of scope: ArcGIS Pro output parity is not claimed. The oracles are server-observed values on the client-compat seed, as described in docs/golden-eval.md.

Cross-links

• Parent: Epic: low-switching-cost Esri migration — .NET import fidelity, JS codemods, then Python GP honua-release#316; gate: release: certify 2026.1 SDK adoption outcomes — .NET import, Esri apps, Python GP honua-release#325.
• Dependents: fix(python-migration): preserve attestation and reject incomplete GP execution as success honua-migrate#143, feat(honua-gp): implement stubs backed by existing server processes and regenerate the compatibility matrix #196, ephemeral-server-smoke blocks commits on a moving honua-server baseline (nightly-aot tag + unpinned seed checkout) #201 (via fix(ci): pin ephemeral-server-smoke to the candidate identity, add a moving canary #228), bug(honua-gp): environment configuration and FeatureServer URL resolution break GetCount parity #205.
• GP output binding: fix(honua-gp): bind Result outputs to real artifacts and reject unsupported parameter semantics #226.

Closes #202