fix(agent): recover inline reasoning stripped by the think scrubber

[SayHell0W0rld]

Summary

Fixes #89647: providers that inline reasoning (MiniMax-M3 streams <think>…</think> inside content) never send a reasoning_content delta, so the desktop reasoning pane stays dead even though reasoning happened.

Approach

• StreamingThinkScrubber (agent/think_scrubber.py, [Bug] MiniMax reasoning content leaks to CLI output even with reasoning_effort: none #17924) already strips inline think blocks from streamed content but discarded the text. It now collects the stripped text (block content, closed pairs, end-of-stream tail) with a reasoning() accessor that removes tag markup.
• _relay_final_response (agent/chat_completion_helpers.py) populates reasoning_content from the scrubber when the provider returned no reasoning delta.

Relation to #43827 / #43836

• This fixes the structured-field extraction side of the MiniMax-M3 reasoning family. The English <think> blocks are already suppressed from visible output by the scrubber ([Bug] MiniMax reasoning content leaks to CLI output even with reasoning_effort: none #17924) — this PR recovers them into reasoning_content so the desktop reasoning pane populates.
• The Chinese reasoning tags (思考/反思/推理/推敲) leak reported in [Bug]: Chinese reasoning tags ( 思考 / 反思 / 推理 / 推敲) from MiniMax-M3 leak to user output in streaming + final response #43827 (still open; fix(agent): add Chinese reasoning tags to think-block filter for MiniMax M3 #43836 closed without merge) is a separate symptom in the same family: those tags are not in the scrubber's tag set, so they are neither suppressed nor collected here. Out of scope for this PR.

Test Plan

• 5 new tests in TestReasoningCollection (tests/agent/test_think_scrubber.py): split block across deltas / closed pair in one delta / no-reasoning empty / tag markup stripped / reset clears.
• Full file: 25/25 pass. ruff check clean.

Risk / Exclusions

• Scrubber behavior for visible output is unchanged (collection only); reasoning() is additive.
• Only fills reasoning_content when the provider gave none — native reasoning deltas still take precedence.
• No changes to streaming delivery, prompts, the content scrubber, or the Chinese-tag path ([Bug]: Chinese reasoning tags ( 思考 / 反思 / 推理 / 推敲) from MiniMax-M3 leak to user output in streaming + final response #43827).

References #89647

---

Update (2026-08-29): This PR now also carries the credential-pool heal fix (formerly #95351, closed). Both changes edit _relay_final_response's success path, so they were merged into one branch to avoid a merge conflict between overlapping edits of the same function:

1. Recover inline reasoning stripped by the think scrubber ([Bug] Hermes Desktop reasoning pane dead for MiniMax-M3 (Plus subscription) — reasoning inline in content, never extracted to reasoning_content #89647)
2. Heal credential-pool last_status after a successful served request (xai-oauth: usage/limits visibility parity with openai-codex (fetch_usage), and heal credential_pool last_status on served requests #95166)

Combined branch: fix/minimax-reasoning-extraction (2 commits). 87/87 tests pass.

[Enough1122]

AI code review — automated review for reference; please use your judgment.

1. agent/think_scrubber.py:reasoning() — Nit: parts are joined with "\n", so a sentence streamed across deltas ("Let me check" + " their config") gains a newline where none existed ("Let me check\ntheir config"). Harmless for model consumption, mildly lossy if a UI renders this as the model's literal reasoning prose. Suggestion: join with "" and let the natural delta boundaries carry whitespace (the scrubber already suppresses nothing but tags), or document the reflow.

2. Overall — Positive: the collection hooks cover all four suppression sites (in-flight block buffer, block-close content, single-delta closed pairs, unterminated flush), reasoning() strips tag markup including case variants, and the relay fallback engages only when no genuine reasoning deltas arrived — so providers with native reasoning streams are untouched. Reset-clears-collected is tested, which matters because a stale turn's reasoning leaking into the next turn's field would be its own bug. Clean fix for the MiniMax-M3 inline-think class of providers.

[SayHell0W0rld]

Requesting review — this PR is a small, self-contained fix with passing tests (full file green, ruff/footgun clean). Happy to address any feedback.