feat(text): align defaults with current M3/Messages API contract

[raylanlin]

Follow-up to #172 (closed). Implements the remaining feedback from @NianJiuZst's review comment against the current Messages API contract.

Background

The M3 default model change from #172 landed in upstream v1.0.18. This PR addresses the other items in NianJiuZst's review that didn't make it in:

1. max_tokens was kept at 4096; the official M3 recommendation is 131072 (128K), other models 65536 (64K).
2. Thinking parameter was not added; Messages API says thinking is disabled by default when omitted, so the CLI must document and expose that as opt-in.
3. Help text documented temperature as (0.0, 1.0]; the M3 Messages API uses [0, 2] with default 1.
4. --top-p had no default documented; M3 Messages API contract default is 0.95.

The stacked quota commit (carrying #168) is removed; rebased cleanly on current main.

Changes

• Model-aware max_tokens (src/commands/text/chat.ts, src/commands/text/repl.ts, src/sdk/text/index.ts, src/utils/model-defaults.ts): 131072 for M3, 65536 for others. --max-tokens flag always overrides. New shared helper resolveMaxTokens() keeps chat, repl, and the SDK in lockstep.
• Thinking parameter (src/types/api.ts): new optional thinking?: {type: 'enabled' | 'disabled' | 'adaptive'} field. New --thinking <mode> CLI flag in both commands. Default: omitted (thinking disabled per API contract). Unknown values throw CLIError (exit code 2 / USAGE) before any network call.
• Temperature [0, 2] default 1 (src/commands/text/chat.ts, src/commands/text/repl.ts, src/utils/model-defaults.ts): help text updated; new resolveTemperature() helper rejects out-of-range and non-numeric values up-front. We do not auto-send 1 when the flag is omitted — backward compat preserved.
• Top-p default 0.95 (src/commands/text/chat.ts, src/commands/text/repl.ts): help text updated to mention 0.95 as the Messages API recommendation. Field still omitted from request body when flag is not passed — no behavior change for existing users.
• Tests (test/utils/model-defaults.test.ts, test/sdk/text.test.ts, test/commands/text/chat.test.ts, test/commands/text/repl.test.ts): focused coverage for each of the above. Coverage spans three altitudes — pure helper unit, SDK round-trip via mock server, and full command execute() via --dry-run.

Quality gates

• typecheck: clean (npx tsc --noEmit exit 0)
• tests: 448 pass / 0 fail / 808 expect() calls (bun test; baseline 403/0/735 → +45 tests)
• build: clean (bun run build exit 0; dist/mmx.mjs + dist/sdk.mjs produced)

Notes

• max_tokens=131072 for M3 follows the official Messages API recommendation. The 524288 (512K) maximum is documented in the API contract but not enforced as a default — explicit --max-tokens overrides remain the user escape hatch (e.g. for --max-tokens 8192).
• Thinking default of "omitted" matches the documented API behavior (thinking disabled by default for M3). If the project later wants a different default, that should be a separate discussion with reason.
• --top-p 0.95 is documented in help but not auto-sent — no behavior change for existing users; new users see the recommended default in --help.
• The resolveMaxTokens, resolveThinkingMode, and resolveTemperature helpers live in a new src/utils/model-defaults.ts so chat/repl/SDK stay in lockstep — preventing the kind of drift that motivated NianJiuZst's review in the first place.

cc @NianJiuZst — thank you for the detailed review on #172.

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}

[raylanlin]

Self-review fix after cross-checking against the current Messages API docs:

The Messages API documents top_p as model-aware:

top_p   number<double>  default:0.95
Nucleus sampling parameter. Range [0, 1]. Default is 0.95 for MiniMax-M3 and 0.9 for M2.x models.

The first push of this PR hardcoded 0.95 in both the TOP_P_DEFAULT export and the --top-p help text, which misrepresented the M2.x default. Replacement (86262ca):

• src/utils/model-defaults.ts: TOP_P_DEFAULT constant → resolveTopPDefault(model) function (M3 → 0.95, anything else → 0.9 as the safe M2.x fallback).
• src/commands/text/chat.ts + src/commands/text/repl.ts: --top-p help text now reflects the per-model default ((default: 0.95 for M3, 0.9 for M2.x per Messages API)).
• test/utils/model-defaults.test.ts: replaced the single TOP_P_DEFAULT assertion with a model-aware resolveTopPDefault describe block (4 cases).

Behavior unchanged: top_p is still only sent on the wire when the user explicitly passes --top-p. The fix is help-text accuracy / contract alignment only — no request-body change for existing users.

Quality gates after the fix: npx tsc --noEmit exit 0, bun test 451 pass / 0 fail / 811 expects (+3 net vs prior push), bun run build clean. Verified the docs at https://platform.minimax.io/docs/api-reference/text-chat-anthropic (which is the page @NianJiuZst cited in the #172 review).

cc @NianJiuZst — only noticed this when I went back to fetch the actual docs after the first push landed. Sorry the first pass oversimplified point 3.