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

[mikemcdougall]

Stacked on #228 (needed so ephemeral-server-smoke is green against a working baseline instead of the currently-broken nightly-aot pin) — the diff will shrink to just this PR's commit once #228 merges and this branch is rebased.

Problem

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). An un-oracled supported script graded as a silent "unblessed" pass, so a response-parsing regression in schema introspection (Describe/ListFields) or the write-cursor round trip (InsertCursor/UpdateCursor) could sail through this lane undetected.

What I found while blessing them

Blessing the two write-cursor scripts against a fresh seed didn't just fill in a missing value — it surfaced a real bug. honua_gp.da._values_for_row / _extract_oid only recognized a feature's .attributes, but the live SDK's honua_sdk.models.QueryFeature (what Source.iter_features actually yields against a real server) exposes GeoJSON-shaped .properties / .id instead — there is no .attributes. Every live SearchCursor/UpdateCursor field read (OID included) silently degraded to None. Stub-mode CI never caught this because the eval stub's _StubFeature happens to define .attributes.

Fixed both helpers to read the real feature shape, preferring the SDK's already-resolved .id for OID@ (the seed's object-id field is the lower-case objectid, which never matched the existing OBJECTID/oid/FID key list either). Added two regression tests against the actual QueryFeature type (verified they fail against the pre-fix code, restored, verified they pass with the fix). Re-blessed against a fresh stack: update_cursor_close_status now genuinely proves an update round-trip (update_count: 1) instead of masking the bug behind a vacuous update_count: 0.

Describe/ListFields (schema fetch, a different code path) and InsertCursor (write-only) were already correct — those simply needed a schema / apply-edits fingerprint oracle.

Non-capturable case: make_table_view

MakeTableView/MakeFeatureLayer resolve to honua_gp.management._make_layer_handler, which only registers a client-side session alias — no honua-server HTTP round trip at all. There's no response to fingerprint. Documented in a new eval/UNBLESSED_ALLOWLIST.json with the reason, per the issue's acceptance criteria ("say so explicitly rather than leaving it implicit").

Anti-regrowth gate

run_eval.py now fails a live-mode supported script that has no response oracle and no UNBLESSED_ALLOWLIST.json entry, instead of passing it as "unblessed". Verified by emptying the allow-list and confirming make_table_view.py fails with the expected reason, then restoring it.

Acceptance criteria (from #202)

• Decided per script whether a response oracle is capturable (8 capturable, 1 genuinely not, documented).
• Blessed the capturable ones against a fresh client-compat seed and committed the oracles.
• Made the remaining unblessed script explicit and non-silent via UNBLESSED_ALLOWLIST.json + a gate that fails a new/unblessed supported script.

Test plan

• python3 -m pytest packages/honua-gp/tests -q: 197 passed (added regression tests confirmed to fail pre-fix, pass post-fix).
• python3 -m honua_gp._cli matrix --check (no drift).
• Stub-mode run_eval.py: 50/50.
• Live-mode run_eval.py (HONUA_GP_EVAL_USE_STUB=0) against three independent fresh seeded stacks: 50/50 each, confirming the blessed oracles are deterministic.
• Confirmed the anti-regrowth gate fails make_table_view.py when its allow-list entry is removed, with the documented reason.

Closes #202

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[chatgpt-codex-connector]

You have reached your Codex usage limits for security reviews. Please try again later.