refactor(gateway): share error-text blob between not_found classifiers

[hashbender]

Follow-up to NousResearch#55780 / NousResearch#56225 (dead-target not_found blast-radius fix, merged).

Why

classify_send_error and is_chat_level_not_found each build a lowercased error blob to substring-match against — but divergently: classify_send_error appends the exception class name (exc.__class__.__name__), while is_chat_level_not_found appends only str(exc). A caller passing exc= to both could get inconsistent answers on the same failure. The two functions' signatures also had swapped positional order ((exc, error_text) vs (error_text, exc)), an easy footgun.

Surfaced by the hermes-pr-review Phase 2c structured review of NousResearch#56225; the merge landed the base fix before this follow-up could ride along, so it's a standalone follow-up.

Changes

• Extract _error_blob(exc, error_text) as the single source of truth both classifiers use (includes str(exc) when non-empty + the class name; skips an empty str(exc) so no stray leading space).
• Align is_chat_level_not_found to (exc, error_text), matching classify_send_error; update the sole caller (gateway/delivery.py) and the three tests to keyword form.
• Add a regression guard asserting _error_blob keeps the class name (mutation-verified: removing the class-name append fails the test).

Tests

pytest tests/gateway/test_dead_targets.py -q
26 passed

No behavior change to the classifier's current call sites (they pass error_text= only).

---

Mirror-of: NousResearch#56241
NousResearch#56241

[tenki-reviewer]

Review Complete
No issues found!

Risk: 🟢 Low (12/100) — no findings · 90 LOC across 3 files

---

Minor refactoring in the gateway delivery pipeline: argument ordering change in BasePlatformAdapter and corresponding callers, with a new is_chat_level_not_found helper on DeliveryTarget. No bugs or regressions detected.

Files Reviewed (3 files)

gateway/delivery.py
gateway/platforms/base.py
tests/gateway/test_dead_targets.py