fix(bindings): sanitize non-JSON-native types at emit() FFI boundary

[Ken Chau (kenotron-ms)]

Summary

• Fixes #228: Non-JSON-native Python types (Decimal, datetime, custom objects) no longer crash emit() and related registry methods when sent as event payloads
• Added json_dumps_safe() DRY helper in helpers.rs that threads default=str through the FFI boundary — all non-JSON types now serialize via their str() representation
• Incidental fix: resolves provider-anthropic cost_usd Decimal serialization issue (related to feat: add --local-source to E2E smoke test + fix input_schema test #52 fix)

Sites Fixed

• bindings/python/src/helpers.rs — new DRY helper
• bindings/python/src/hooks.rs — emit(), set_default_fields(), emit_and_collect()
• bindings/python/src/bridges.rs — PyHookHandlerBridge::handle() result path
• python/amplifier_core/loader_grpc.py — GrpcToolBridge._serialize_input()

Test Coverage

• 5 new tests in bindings/python/tests/test_emit_json_sanitize.py
• 1 new test in bindings/python/tests/test_loader_grpc.py
• Full suite: 1016 passed, 1 skipped, 0 failed

Design doc: docs/plans/2026-05-08-emit-json-sanitize-design.md

---

Generated with Amplifier

[Ken Chau (kenotron-ms)]

DTU Validation — Pre-fix provider-anthropic + fixed amplifier-core

To validate that the default=str safety net at the FFI boundary is sufficient on its own (independent of any provider-side fix), a Digital Twin Universe was run with the following deliberately adversarial stack:

Component	Version / State
amplifier-core	fix/emit-json-sanitize-228 — this PR (1.5.2 wheel, built inside DTU)
provider-anthropic	c07131df6a123809d3806b72599732ab3fa71197 — base of PR #53, before the provider-side str() conversion landed
amplifier CLI + foundation	upstream main

The provider at that commit passes cost_usd (a Decimal) directly into the llm:response event dict without converting it — this is the exact code path that caused the original crash. The question being tested: does the Core fix catch it even when the provider doesn't?

Results

[PASS] amplifier, version 2026.05.08-a0e1627 (core 1.5.2) — branch wheel confirmed loaded
[PASS] provider-anthropic @ c07131df confirmed installed (settings.yaml source pin)
[PASS] ANTHROPIC_API_KEY passed through
[PASS] amplifier run "Say exactly: amplifier-tester-ok"
         Tokens:  57,377 in / 61 out / 57,438 total
         Cost:    $0.22
         Reply:   amplifier-tester-ok
         Exit:    0
[PASS] No TypeError, no JSON serialization crash anywhere in the run

Why the cost stamp is the proof

The $0.22 line is rendered by the cost-tracking hook that consumes the llm:response event. For that to print, the event payload had to pass through emit() at the Python→Rust FFI boundary — the exact call site that this PR hardens with json_dumps_safe(). The session exited 0, confirming the boundary now tolerates non-JSON-native types regardless of what the provider passes.

Also confirmed

• Downstream repo check: amplifier, amplifier-app-cli, amplifier-foundation — no [tool.uv.sources] git overrides for amplifier-core ✅
• E2E smoke test (from scripts/e2e-smoke-test.sh): PASSED — manylinux wheel built, pristine-import preflight passed, core 1.5.2 confirmed, no crashes ✅