Improve MiniMax/Kimi provider ergonomics + stabilize slides/progress output

[spirosrap]

Improve MiniMax/Kimi provider ergonomics + stabilize slides/progress output

Summary

This PR started with provider/model ergonomics, and now also includes a substantial hardening pass for slides-mode output quality based on real user runs.

1) MiniMax/Kimi provider ergonomics

• Add first-class --model kimi alias (defaulting to openai/kimi-k2.5)
• Add --model kimi/<model>
• Keep/clarify MiniMax alias flow (--model minimax, --model minimax/<model>)
• Add Kimi env support:
  • KIMI_API_KEY (primary)
  • MOONSHOT_API_KEY (alias)
  • KIMI_BASE_URL (primary)
  • MOONSHOT_BASE_URL (alias)
  • default base URL: https://api.moonshot.ai/v1
• Wire through model parsing, auto-selection, env/runtime context, daemon, URL/asset flows, and help/docs

2) Kimi runtime compatibility fix

• For openai/kimi..., force effective temperature=1 to avoid provider runtime errors (only 1 is allowed).

3) Slides/progress output stabilization

• Fix spinner/progress gate races with stdout
• Improve slide marker/chunk handling to avoid dropped interleaved text
• Keep deterministic non-inline terminal behavior

4) Slide summary quality hardening (post-feedback)

• Reduce mid-sentence truncation artifacts
• Prevent oversized raw transcript blob injection into slide blocks
• Compact oversized slide blocks while preserving sentence-complete output
• Strengthen slide prompt instructions for concise third-person summaries
• Suppress transcript-style fallback injection (dialog/CTA/disfluency tails)
• Normalize conversational tails (e.g. even like, ..., right?, in a way of like)
• Strip single-bullet monologue formatting and clean pronoun rewrite artifacts (the speaker the speaker)
• Strip leading disfluencies in transcript-like slide tails (Um/uh) and normalize them into neutral prose

Why

Users reported:

• cumbersome provider setup for non-default OpenAI-compatible endpoints
• Kimi runtime failures with temperature handling
• slides mode truncation/misalignment
• transcript-like raw speech leaking into final slide paragraphs

This PR addresses all of the above in one branch.

User-facing behavior changes

• New/expanded model aliases:
  • summarize "<url>" --model kimi
  • summarize "<url>" --model kimi/<model-id>
  • summarize "<url>" --model minimax
  • summarize "<url>" --model minimax/<model-id>
• Improved slides-mode robustness and consistency on long videos and streamed chunk boundaries

Docs

• README.md
• docs/llm.md
• docs/config.md
• CLI help text (src/run/help.ts)

Tests / validation

Added/updated coverage for:

• model parsing/spec + auto-selection (kimi, aliases, MiniMax)
• config env mapping and daemon key selection
• Kimi temperature handling
• slides marker parsing, fallback segmentation, truncation regression cases
• transcript-style tail normalization in slide blocks
• progress gate stream stability

Representative test runs:

npm exec --yes pnpm@latest -- -s test \
  tests/model-spec.test.ts \
  tests/model-auto.test.ts \
  tests/config.env.test.ts \
  tests/daemon.agent.test.ts \
  tests/llm.generate-text.test.ts \
  tests/slides-output.stream-handler.test.ts \
  tests/slides-text.utils.test.ts \
  tests/slides.output.parse.test.ts \
  tests/progress-gate.test.ts \
  tests/summary-sanitize.test.ts \
  tests/prompts.link-summary.slides.test.ts

npm exec --yes pnpm@latest -- -s build

Commit history note

• The high commit count on this branch is intentional.
• Most follow-up commits are focused, incremental fixes for MiniMax slide rendering/output quality regressions found in real terminal/video runs (especially final-slide transcript-tail coherence and truncation behavior).
• Each fix was validated with targeted tests and rebuilt to keep behavior stable while iterating quickly.