Fix OpenAI strict-mode schema for /api/memory/extract + /api/decide

[wdwd720]

Stacked on PR #4 (cursor/llm-memory-and-reasoning-c3fa) so this PR's diff is just the schema fix.

The bug

OpenAI's strict structured outputs require every object schema to list every property key in required (optional fields stay required but nullable). Our nested items[] schema in MEMORY_EXTRACTION_SCHEMA only listed 5 of its 12 keys in required, so the API rejected the request with:

Invalid schema for response_format 'hearer_response':
  In context=('properties', 'items', 'items'), 'required' is required to be supplied

Result: /api/memory/extract silently fell back to mode: "rule_based", even with enabled=true configured=true provider=openai_compatible.

The fix

server/src/llm/llmSchemas.ts

• New strictObject(properties) helper. It auto-derives required from Object.keys(properties) so it's impossible to forget a field. Sets type: "object", additionalProperties: false, required, and properties.
• MEMORY_EXTRACTION_SCHEMA rebuilt with strictObject(). The inner item schema now lists all 12 keys (kind, label, triggerDescription, actionType, actionLabel, person, deadlineText, locationLabel, priority, confidence, save, privacyNote) in required. Optional fields stay required but use nullable union types (type: ["string", "null"], or enum: [..., null] for priority).
• CUE_REASONING_SCHEMA rebuilt the same way for safety.
• New exported assertStrictOpenAiSchema(schema, path?) walks a schema and throws on the first violation (missing properties / required / additionalProperties, mismatch between required and properties keys, recurses into array items).

server/src/llm/openAiCompatibleProvider.ts

• chatJson now takes a schemaName, so each request gets its own name in the OpenAI payload: "hearer_memory_extraction" / "hearer_cue_reasoning".

• On a non-2xx response we parse OpenAI's error.message + error.code and surface a compact line in the server log:

  [hearer-llm] provider HTTP 400 on schema='hearer_memory_extraction': invalid_request_error: …
  

  We never log the API key (it's in the Authorization header) and we never log the user transcript.

Tests

server/src/__tests__/openAiSchemaStrictness.test.ts (new):

• MEMORY_EXTRACTION_SCHEMA passes assertStrictOpenAiSchema.
• CUE_REASONING_SCHEMA passes assertStrictOpenAiSchema.
• The nested memory item schema's required matches its properties keys (every one of the 12 fields).
• The assertion throws when required is missing a property.
• The assertion throws when additionalProperties is missing.
• The assertion recurses into array item schemas.

README

Smoke commands added with the expected mode: "llm", fallbackUsed: false response.

Results

• npm run build → success.
• npm test → 71 / 71 passing across 19 suites (was 65 / 18). All prior tests still pass.

After merging this stack

curl -s http://localhost:8788/api/llm/status
# { "enabled": true, "configured": true, "provider": "openai_compatible", "model": "gpt-4.1-mini", ... }

curl -s -X POST http://localhost:8788/api/memory/extract \
  -H 'content-type: application/json' \
  -d '{"text":"Usually after dinner I cook, so if you hear beeping remind me to check the stove."}'
# {
#   "mode": "llm",
#   "fallbackUsed": false,
#   "persisted": { "routines": [ { "name": "Cooking timer safety cue", "actionLabel": "check stove", ... } ] },
#   ...
# }