feat(cognitive): improve historian quality and stream LLM metrics - #92

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@docs/message-batching.md`:
- Line 46: Define one consistent observation provenance contract across
docs/message-batching.md:46-46, res/prompts/undefined.xml:240-252 and :961-965,
and res/prompts/undefined_nagaagent.xml:240-252 and :1020-1024. Decide whether
verified assistant-action outcomes are valid observation sources, document that
decision, and update each end.observations and memory-management prompt to
enforce the same rule; preserve the requirement that observations are otherwise
limited to substantive facts from the current input batch and use an empty array
when none exist.

In `@res/prompts/historian_profile_merge.md`:
- Around line 118-135: Update the fenced code block in the positive example to
specify the text language on its opening fence, resolving the MD040 markdownlint
violation while preserving the example content.

In `@res/prompts/historian_rewrite.md`:
- Around line 7-21: Update the examples in the “轻量独立事实” section to use the
required format “QQ号123（昵称张三）” instead of “张三(123)”. Also revise the name-format
rule around line 25 to require cognitive observations to use
“QQ号<user_id>（昵称<name>）”, preserving the explicit numeric identifier.

In `@src/Undefined/ai/llm/streaming.py`:
- Around line 246-266: Update responses_event_marks_ttft so
response.function_call_arguments.delta events return
bool(stringify_stream_delta(event.get("delta"))) instead of unconditionally
marking TTFT. Add a regression test covering an event with an empty string delta
and assert it is not treated as TTFT.

In `@src/Undefined/cognitive/historian/worker.py`:
- Around line 680-705: Update the timestamp setup around now_local_dt,
now_utc_dt, and timezone_label to resolve job["timezone"] with ZoneInfo, using a
safe fallback for invalid or missing values. Derive local and UTC timestamps
from the same current instant, ensure the prompt’s now_local uses the resolved
job timezone, and emit future updated_at values with an explicit offset; apply
the same correction to the corresponding logic near the profile-merge handling
at lines 732-736.

In `@src/Undefined/skills/tools/end/config.json`:
- Line 16: Require user-centered cognitive-memory observations to identify users
as QQ号<sender_id>（昵称<name>）, updating the observations description in
src/Undefined/skills/tools/end/config.json:16-16 and the function description at
src/Undefined/skills/tools/end/config.json:5-5. Mirror this stable-ID
requirement in docs/cognitive-memory.md:71-71 and docs/cognitive-memory.md:8-8,
and align src/Undefined/skills/tools/end/README.md:7-7. Extend
tests/test_system_prompt_constraints.py:448-455 and
tests/test_system_prompt_constraints.py:492-498 to assert the rule in both
prompt and end-tool schema contracts.

Apply the same fix in `@docs/cognitive-memory.md` at line 8.

Apply the same fix in `@tests/test_system_prompt_constraints.py` around lines 448
- 455.

Apply the same fix in `@src/Undefined/skills/tools/end/README.md` at line 7.

In `@src/Undefined/token_usage_storage.py`:
- Around line 83-89: Update to_optional_float to return None when the parsed
float is non-finite or negative, while preserving None for null and invalid
inputs and valid non-negative finite values. Add tests covering "NaN",
"Infinity", and negative values.

In `@tests/test_cognitive_historian.py`:
- Around line 375-385: Update the merge template contract test assertions around
the existing now_local check to also require the {now_utc} placeholder and
verify the generated prompt includes a valid UTC timestamp value. Preserve the
existing assertions and add coverage that would fail if the worker stops
supplying or rendering now_utc.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@tests/test_cognitive_historian.py`:
- Around line 423-428: Update test_now_in_job_timezone_invalid_falls_back to
compare now_local’s UTC offset and timezone name against
datetime.now().astimezone(), asserting the invalid timezone falls back to the
system-local timezone rather than merely checking the label differs. Retain the
existing UTC and consistency assertions.

---

Outside diff comments:
In `@tests/test_cognitive_historian.py`:
- Around line 403-411: Add cases to test_extract_frontmatter_updated_at for
malformed YAML and valid YAML whose top-level value is not a dictionary,
asserting both return an empty string. Keep the existing timestamp,
missing-field, and absent-frontmatter assertions unchanged.