pi-voicekit 0.2.3 → 0.3.1

package/extensions/voice/post-process-queue.ts

@@ -0,0 +1,532 @@
+/**
+ * Segmented transcript post-processing: a bounded, fail-open polish queue.
+ *
+ * Pure and dependency-injected like ./post-process: no Pi types, no TUI, no filesystem
+ * and no network. The model call arrives as `call`, so ordering, the concurrency cap,
+ * the per-segment timeout, the single retry and per-segment isolation are all provable offline.
+ *
+ * The cap counts real calls, not segment promises: a slot is released when the injected
+ * transport settles, so a caller that ignores the abort signal keeps its slot after its
+ * segment timed out instead of letting a hung endpoint pile up calls past the cap.
+ *
+ * Spec: docs/superpowers/specs/2026-09-26-polish-pipeline-design.md §4.2, §4.3, §4.4
+ * (a local design record, not part of the published package)
+ */
+
+import {
+	polishSamplingOptions,
+	polishTranscript,
+	type AssistantLike,
+	type PolishRequest,
+	type PolishResult,
+} from "./post-process";
+import { assembleContext, DEFAULT_CONTEXT_LIMITS, type ContextLimits, type EntryLike } from "./post-process-context";
+
+/**
+ * Start at most this many segment calls at once. A long dictation must not open one request
+ * per segment, and wall-clock must stay close to the slowest segment rather than their sum
+ * (spec §4.2, open question 1).
+ */
+export const POLISH_QUEUE_CONCURRENCY = 3;
+
+/** The extra request fields `polishSamplingOptions` decided for one segment. */
+export interface PolishSampling {
+	samplingParams?: { reasoning_effort: string };
+}
+
+/**
+ * The polish request as the injected caller sees it: the prompt plus this segment's sampling
+ * decision. The caller forwards `samplingParams` to its transport options; it is absent when
+ * the model has no thinking to gate.
+ */
+export type QueuePolishRequest = PolishRequest & PolishSampling;
+
+/** The model call, injected. The queue decides `samplingParams` per segment. */
+export type PolishQueueCaller = (request: QueuePolishRequest, signal: AbortSignal) => Promise<AssistantLike>;
+
+export interface SegmentOutcome {
+	index: number;
+	status: "applied" | "fallback";
+	/** What this segment contributes: the rewrite when applied, its own raw text otherwise. */
+	text: string;
+	reason?: string;
+	retried: boolean;
+	latencyMs: number;
+}
+
+export interface QueueResult {
+	/** The ordered join of every segment outcome (spec §4.2 stitching rule). */
+	text: string;
+	segments: SegmentOutcome[];
+	polished: number;
+	failed: number;
+	retried: number;
+}
+
+export interface PolishQueue {
+	/** Enqueue one recogniser segment. Empty text is skipped; work starts immediately. */
+	push(index: number, raw: string): void;
+	/** Barrier for one dictation: resolves once every segment pushed before this call settled. */
+	finish(): Promise<QueueResult>;
+}
+
+export interface PolishQueueOptions {
+	call: PolishQueueCaller;
+	/** Per-segment timeout, the same knob as the single-call path (`postProcessTimeoutMs`). */
+	timeoutMs: number;
+	/** Session context entries; attached to segment 0 only (spec §4.3). */
+	entries?: readonly EntryLike[];
+	limits?: ContextLimits;
+	/** The model whose thinking policy gates each segment; absent or non-reasoning keeps thinking on. */
+	model?: { reasoning?: boolean } | undefined | null;
+	/** Defaults to POLISH_QUEUE_CONCURRENCY; clamped to at least one. */
+	concurrency?: number;
+	/**
+	 * One pass id for the whole queue. Checked before each attempt is scheduled (first try and
+	 * retry), so an invalidated pass stops sending new requests instead of merely discarding
+	 * the answers; `polishTranscript` checks it again before and after every call.
+	 */
+	isCurrent?: () => boolean;
+	/** Injected clock, so tests can pin latency without waiting. */
+	now?: () => number;
+	debug?: (reason: string, data?: Record<string, unknown>) => void;
+}
+
+interface SegmentJob {
+	index: number;
+	raw: string;
+}
+
+/** One real call, which owns a concurrency slot until the transport settles - timeout or not. */
+interface LiveCall {
+	settled: boolean;
+	orphaned: boolean;
+}
+
+/** One attempt: its result, whether a request really went out, or that no request was sent. */
+interface AttemptRun {
+	issued: boolean;
+	denied: boolean;
+	/** The pass lost ownership before this attempt could send anything. */
+	stale: boolean;
+	result: PolishResult | null;
+	error?: unknown;
+}
+
+/**
+ * Stitch segment texts into one transcript. Two adjacent CJK characters take no separator,
+ * so Chinese stays "这是第一段这是第二段" while English keeps "first part second part"; every
+ * other boundary takes a single space, which is what zh-en dictation reads as. Empty parts
+ * are dropped so a silent segment cannot inject a stray space, and each part is trimmed at
+ * its edges only.
+ */
+export function joinSegments(parts: readonly string[]): string {
+	let joined = "";
+	for (const part of parts) {
+		const text = part.trim();
+		if (!text) continue;
+		if (joined && needsSeparator(joined, text)) joined += " ";
+		joined += text;
+	}
+	return joined;
+}
+
+function needsSeparator(left: string, right: string): boolean {
+	return !(isCjk(lastCodePoint(left)) && isCjk(firstCodePoint(right)));
+}
+
+function firstCodePoint(text: string): number {
+	return text.codePointAt(0) ?? -1;
+}
+
+/** The code point at the end of `text`, surrogate pairs included. */
+function lastCodePoint(text: string): number {
+	const last = text.charCodeAt(text.length - 1);
+	const previous = text.charCodeAt(text.length - 2);
+	if (last >= 0xdc00 && last <= 0xdfff && previous >= 0xd800 && previous <= 0xdbff) {
+		return (previous - 0xd800) * 0x400 + (last - 0xdc00) + 0x10000;
+	}
+	return last;
+}
+
+/**
+ * CJK for the stitching rule: ideographs and their extensions, kana, Bopomofo, CJK
+ * punctuation and fullwidth forms. Hangul is left out on purpose — Korean separates words
+ * with spaces.
+ */
+function isCjk(codePoint: number): boolean {
+	return (
+		(codePoint >= 0x2e80 && codePoint <= 0x2eff) || // CJK Radicals Supplement
+		(codePoint >= 0x3000 && codePoint <= 0x303f) || // CJK Symbols and Punctuation
+		(codePoint >= 0x3040 && codePoint <= 0x30ff) || // Hiragana and Katakana
+		(codePoint >= 0x3100 && codePoint <= 0x312f) || // Bopomofo
+		(codePoint >= 0x3190 && codePoint <= 0x319f) || // Kanbun
+		(codePoint >= 0x31c0 && codePoint <= 0x31ef) || // CJK Strokes
+		(codePoint >= 0x3400 && codePoint <= 0x4dbf) || // CJK Unified Ideographs Extension A
+		(codePoint >= 0x4e00 && codePoint <= 0x9fff) || // CJK Unified Ideographs
+		(codePoint >= 0xf900 && codePoint <= 0xfaff) || // CJK Compatibility Ideographs
+		(codePoint >= 0xfe30 && codePoint <= 0xfe4f) || // CJK Compatibility Forms
+		(codePoint >= 0xff01 && codePoint <= 0xff60) || // Fullwidth forms
+		(codePoint >= 0x20000 && codePoint <= 0x3ffff) // Extensions B and beyond
+	);
+}
+
+/** Same shape as `resolveTurns` in post-process-context: floor locally, never let 0 through. */
+function resolveConcurrency(value: number | undefined): number {
+	if (value === undefined || !Number.isFinite(value)) return POLISH_QUEUE_CONCURRENCY;
+	return Math.max(1, Math.floor(value));
+}
+
+/**
+ * One polish queue per dictation. Every segment runs `polishTranscript` with the same
+ * guardrails as the single-call path, so nothing here re-implements the prompt, the output
+ * validation or the fail-open fallback.
+ */
+export function createPolishQueue(options: PolishQueueOptions): PolishQueue {
+	const concurrency = resolveConcurrency(options.concurrency);
+	const limits = options.limits ?? DEFAULT_CONTEXT_LIMITS;
+	const clock = (): number => {
+		try {
+			return (options.now ?? Date.now)();
+		} catch {
+			// A broken injected clock must not cost the user their dictation.
+			return Date.now();
+		}
+	};
+	/** Raw text by index: the join order, and the reference turn for the next segment. */
+	const raws = new Map<number, string>();
+	const outcomes = new Map<number, SegmentOutcome>();
+	/**
+	 * Real calls in flight, and how many of them outlived their segment. A slot is released when
+	 * the transport settles, never when the segment deadline fires: an injected caller that
+	 * ignores the abort signal keeps its slot, so a hung endpoint cannot push the live request
+	 * count past `concurrency`. `running` counts segments whose outcome is not decided yet.
+	 */
+	let liveCalls = 0;
+	let orphanedCalls = 0;
+	let running = 0;
+	const slotWaiters: ((granted: boolean) => void)[] = [];
+	let finished: Promise<QueueResult> | null = null;
+	let resolveIdle: (() => void) | null = null;
+
+	function orderedIndexes(): number[] {
+		return [...raws.keys()].sort((left, right) => left - right);
+	}
+
+	function hasFreeSlot(): boolean {
+		return liveCalls < concurrency;
+	}
+
+	/**
+	 * True when no slot can free on its own: every live call has already outlived its deadline,
+	 * so waiting only delays the fallback. Every issued call sits inside a `polishTranscript`
+	 * timeout, so a live call either settles or becomes an orphan within one deadline.
+	 */
+	function capacityStuck(): boolean {
+		return !hasFreeSlot() && orphanedCalls >= liveCalls;
+	}
+
+	/** After a slot frees or an orphan appears: grant while slots last, else deny when stuck. */
+	function wakeWaiters(): void {
+		while (slotWaiters.length > 0) {
+			if (hasFreeSlot()) {
+				liveCalls += 1;
+				slotWaiters.shift()!(true);
+				continue;
+			}
+			if (capacityStuck()) {
+				for (const waiter of slotWaiters.splice(0)) waiter(false);
+			}
+			return;
+		}
+	}
+
+	/** Take a slot right now, so a pushed segment reaches its caller without a microtask delay. */
+	function tryTakeSlot(): boolean {
+		if (!hasFreeSlot()) return false;
+		liveCalls += 1;
+		return true;
+	}
+
+	/** Wait for a slot; false means no live call will ever free one, so the segment falls back. */
+	function waitForSlot(): Promise<boolean> {
+		if (tryTakeSlot()) return Promise.resolve(true);
+		if (capacityStuck()) return Promise.resolve(false);
+		return new Promise<boolean>((resolve) => {
+			slotWaiters.push(resolve);
+		});
+	}
+
+	function releaseSlot(orphaned: boolean): void {
+		if (orphaned) orphanedCalls = Math.max(0, orphanedCalls - 1);
+		liveCalls = Math.max(0, liveCalls - 1);
+		wakeWaiters();
+	}
+
+	/** The attempt is over but its call never settled: the slot stays held until it does. */
+	function markOrphaned(call: LiveCall): void {
+		if (call.orphaned) return;
+		call.orphaned = true;
+		orphanedCalls += 1;
+		wakeWaiters();
+	}
+
+	function resolveIdleNow(): void {
+		if (!resolveIdle || running > 0) return;
+		const resolve = resolveIdle;
+		resolveIdle = null;
+		resolve();
+	}
+
+	function settle(job: SegmentJob, outcome: SegmentOutcome): void {
+		outcomes.set(job.index, outcome);
+		running -= 1;
+		resolveIdleNow();
+	}
+
+	/**
+	 * Spec §4.3: segment 0 carries the session context; a later segment carries the previous
+	 * segment's raw text as a single reference turn instead. The reference turn goes through
+	 * `assembleContext` so the same caps apply to it as to the session context.
+	 */
+	function entriesForSegment(index: number): readonly EntryLike[] {
+		if (index === 0) return options.entries ?? [];
+		const previous = raws.get(index - 1);
+		if (!previous) return [];
+		const reference = assembleContext([{ type: "message", message: { role: "user", content: previous } }], limits);
+		const text = reference.turns
+			.map((turn) => turn.text)
+			.join("\n")
+			.trim();
+		if (!text) return [];
+		return [{ type: "message", message: { role: "user", content: text } }];
+	}
+
+	function debugFor(index: number): ((reason: string, data?: Record<string, unknown>) => void) | undefined {
+		const debug = options.debug;
+		if (!debug) return undefined;
+		return (reason, data) => {
+			try {
+				debug(reason, { ...data, segment: index });
+			} catch {
+				// The debug hook is observational: a throw here must not break fail-open.
+			}
+		};
+	}
+
+	function appliedOutcome(job: SegmentJob, text: string, retried: boolean, started: number): SegmentOutcome {
+		return { index: job.index, status: "applied", text, retried, latencyMs: Math.max(0, clock() - started) };
+	}
+
+	function fallbackOutcome(
+		job: SegmentJob,
+		reason: string | undefined,
+		retried: boolean,
+		started: number
+	): SegmentOutcome {
+		const outcome: SegmentOutcome = {
+			index: job.index,
+			status: "fallback",
+			text: job.raw,
+			retried,
+			latencyMs: Math.max(0, clock() - started),
+		};
+		if (reason !== undefined) outcome.reason = reason;
+		return outcome;
+	}
+
+	function reasonOf(run: AttemptRun): string | undefined {
+		if (run.stale) return "invalidated";
+		if (run.error !== undefined) {
+			return run.error instanceof Error && run.error.message ? run.error.message : "segment-error";
+		}
+		if (run.denied) return "no-capacity";
+		return run.result?.reason;
+	}
+
+	/** True once the pass that owns this queue has been invalidated. A throwing check keeps
+	 * fail-open intact: it must never block a dictation that might still be valid. */
+	function stale(): boolean {
+		try {
+			return options.isCurrent?.() === false;
+		} catch {
+			return false;
+		}
+	}
+
+	/**
+	 * Spec §4.2: one retry, and only for a call that timed out or failed at the transport. A
+	 * guardrail rejection means the rewrite was wrong rather than slow, so repeating it cannot
+	 * help; an attempt that never issued a request has nothing to repeat.
+	 */
+	function isRetryable(run: AttemptRun): boolean {
+		if (!run.issued || !run.result || run.result.status !== "rejected") return false;
+		return run.result.reason === "timeout" || run.result.reason === "call-failed";
+	}
+
+	function unexpectedOutcome(job: SegmentJob, error: unknown, latencyMs: number): SegmentOutcome {
+		return {
+			index: job.index,
+			status: "fallback",
+			text: job.raw,
+			reason: error instanceof Error && error.message ? error.message : "segment-error",
+			retried: false,
+			latencyMs,
+		};
+	}
+
+	/**
+	 * One model attempt. The slot is taken before `polishTranscript` starts its deadline, so a
+	 * segment that waited for capacity still gets its full timeout, and the slot goes back only
+	 * once the transport settles.
+	 */
+	async function runAttempt(job: SegmentJob, forceOff: boolean): Promise<AttemptRun> {
+		// Invalidation is checked before a slot is taken: a cancelled or superseded pass must stop
+		// scheduling requests, not just ignore their answers.
+		if (stale()) return { issued: false, denied: false, stale: true, result: null };
+		if (!tryTakeSlot() && !(await waitForSlot())) return { issued: false, denied: true, stale: false, result: null };
+		// The pass may have died while this segment waited for a slot.
+		if (stale()) {
+			releaseSlot(false);
+			return { issued: false, denied: false, stale: true, result: null };
+		}
+		const call: LiveCall = { settled: false, orphaned: false };
+		let issued = false;
+		try {
+			const sampling = polishSamplingOptions(options.model);
+			const result = await polishTranscript({
+				raw: job.raw,
+				entries: entriesForSegment(job.index),
+				limits,
+				timeoutMs: options.timeoutMs,
+				timestamp: clock(),
+				isCurrent: options.isCurrent,
+				call: async (request, signal) => {
+					issued = true;
+					try {
+						return await options.call({ ...request, ...sampling }, signal);
+					} finally {
+						call.settled = true;
+						releaseSlot(call.orphaned);
+					}
+				},
+				debug: debugFor(job.index),
+			});
+			if (!call.settled) markOrphaned(call);
+			return { issued, denied: false, stale: false, result };
+		} catch (error) {
+			// `polishTranscript` normally reports through its result; this only covers a fault
+			// thrown before it could invoke the caller.
+			if (!issued) releaseSlot(false);
+			else if (!call.settled) markOrphaned(call);
+			return { issued, denied: false, stale: false, result: null, error };
+		}
+	}
+
+	async function runSegment(job: SegmentJob): Promise<SegmentOutcome> {
+		const started = clock();
+		try {
+			const first = await runAttempt(job, false);
+			if (first.result?.status === "applied") return appliedOutcome(job, first.result.text, false, started);
+			const firstReason = reasonOf(first);
+			if (!isRetryable(first)) return fallbackOutcome(job, firstReason, false, started);
+			// Never retry for a pass that lost ownership while the first attempt ran.
+			if (stale()) return fallbackOutcome(job, "invalidated", false, started);
+			const second = await runAttempt(job, true);
+			if (second.result?.status === "applied") return appliedOutcome(job, second.result.text, second.issued, started);
+			// A retry that never reached the transport leaves the first failure as the reason.
+			const reason = second.issued ? (reasonOf(second) ?? firstReason) : firstReason;
+			return fallbackOutcome(job, reason, second.issued, started);
+		} catch (error) {
+			// Fail-open, last resort: this segment keeps its own raw text.
+			return unexpectedOutcome(job, error, Math.max(0, clock() - started));
+		}
+	}
+
+	function buildResult(): QueueResult {
+		const segments: SegmentOutcome[] = [];
+		for (const index of orderedIndexes()) {
+			const outcome = outcomes.get(index);
+			if (outcome) {
+				segments.push(outcome);
+			} else {
+				// A pushed segment always settles before finish() resolves, so this only fires if
+				// an unexpected throw slipped past the per-segment catch. Keep the raw text.
+				segments.push({
+					index,
+					status: "fallback",
+					text: raws.get(index) ?? "",
+					reason: "missing-outcome",
+					retried: false,
+					latencyMs: 0,
+				});
+			}
+		}
+		return {
+			text: joinSegments(segments.map((segment) => segment.text)),
+			segments,
+			polished: segments.filter((segment) => segment.status === "applied").length,
+			failed: segments.filter((segment) => segment.status === "fallback").length,
+			retried: segments.filter((segment) => segment.retried).length,
+		};
+	}
+
+	/** Every segment raw, order kept: the last-resort result if the builder itself faults. */
+	function rawOnlyResult(reason: string): QueueResult {
+		const segments: SegmentOutcome[] = orderedIndexes().map((index) => ({
+			index,
+			status: "fallback" as const,
+			text: raws.get(index) ?? "",
+			reason,
+			retried: false,
+			latencyMs: 0,
+		}));
+		return {
+			text: joinSegments(segments.map((segment) => segment.text)),
+			segments,
+			polished: 0,
+			failed: segments.length,
+			retried: 0,
+		};
+	}
+
+	return {
+		push(index: number, raw: string): void {
+			// finish() is the barrier for one dictation: anything offered afterwards belongs to
+			// a pass that is already being written, so it can only be dropped.
+			if (finished) return;
+			// Silence contributes nothing, exactly as today's segment concatenation skips it.
+			if (!raw.trim()) return;
+			// One segment per index: a repeated push would duplicate the text in the join.
+			if (raws.has(index)) return;
+			raws.set(index, raw);
+			running += 1;
+			const job: SegmentJob = { index, raw };
+			// Every pushed segment starts here; the slot pool, not this call site, bounds requests.
+			void runSegment(job).then(
+				(outcome) => settle(job, outcome),
+				(error: unknown) => settle(job, unexpectedOutcome(job, error, 0))
+			);
+		},
+
+		finish(): Promise<QueueResult> {
+			if (!finished) {
+				finished = new Promise<QueueResult>((resolve) => {
+					// The result builder is the last place a fault could cost the dictation its
+					// text, so a throw still resolves with the ordered raw transcript.
+					resolveIdle = () => {
+						try {
+							resolve(buildResult());
+						} catch {
+							resolve(rawOnlyResult("result-error"));
+						}
+					};
+				});
+				// Nothing pushed: resolve in this tick instead of waiting for a segment that may
+				// never come.
+				resolveIdleNow();
+			}
+			return finished;
+		},
+	};
+}

package/extensions/voice/post-process.ts

@@ -142,7 +142,7 @@ export type EditorRead = string | typeof EDITOR_READ_FAILED;
  * while the pass still owns the flow AND the editor still holds the value the pass
  * snapshotted. Used by the normal path and by the pass's own throw path.
  */
-export function decideApply(input: { tokenCurrent: boolean; editorSnapshot: string; currentEditor: EditorRead }): {
+export function decideApply(input: { tokenCurrent: boolean; editorSnapshot: EditorRead; currentEditor: EditorRead }): {
 	apply: boolean;
 	reason?: string;
 } {
@@ -180,6 +180,12 @@ export async function polishTranscript(input: PolishInput): Promise<PolishResult
 	const context = assembleContext(input.entries, input.limits);
 	const shape = { contextChars: context.characters, truncatedContext: context.truncated };
 	const request = buildPolishRequest(context, input.raw, input.timestamp);
+	// Invalidation stops scheduling, not merely the write: a cancelled or superseded pass must
+	// not send transcript text to the provider at all, so the check in front of the call is as
+	// important as the one after it.
+	if (input.isCurrent && !input.isCurrent()) {
+		return { status: "skipped", text: input.raw, reason: "invalidated", ...shape };
+	}
 	const controller = new AbortController();
 	let timer: ReturnType<typeof setTimeout> | undefined;
 
@@ -218,6 +224,17 @@ export async function polishTranscript(input: PolishInput): Promise<PolishResult
 	}
 }
 
+/**
+ * Per-segment outcome of the pipelined pass, one summary per dictation. Absent for the
+ * single-call path, where there are no recogniser segments to report.
+ */
+export interface PolishAuditSegments {
+	count: number;
+	polished: number;
+	failed: number;
+	retried: number;
+}
+
 /**
  * One durable record of what a pass did, written into the session file by the caller.
  *
@@ -248,6 +265,14 @@ export interface PolishAudit {
 	durationSec?: number;
 	/** Which recogniser produced it, so results are never pooled across backends. */
 	backend?: string;
+	/**
+	 * The local model that produced the text (e.g. "parakeet-v3"), so dictations can be
+	 * grouped by recogniser within the local backend. Absent for cloud backends, which
+	 * carry their own server-side model choice.
+	 */
+	recognizer?: string;
+	/** Per-segment outcome when the segmented queue produced this dictation. */
+	segments?: PolishAuditSegments;
 	/** True when the pass asked the model not to think. */
 	thinkingOff?: boolean;
 	/** The output-token cap the request carried, a cap and not a spend. */
@@ -265,6 +290,8 @@ export function buildPolishAudit(input: {
 	thinkingOff?: boolean;
 	durationSec?: number;
 	backend?: string;
+	recognizer?: string;
+	segments?: PolishAuditSegments;
 	maxTokens?: number;
 	telemetry?: {
 		model?: string;
@@ -290,6 +317,8 @@ export function buildPolishAudit(input: {
 	if (input.maxTokens !== undefined) audit.maxTokens = input.maxTokens;
 	if (input.durationSec !== undefined) audit.durationSec = input.durationSec;
 	if (input.backend !== undefined) audit.backend = input.backend;
+	if (input.recognizer !== undefined) audit.recognizer = input.recognizer;
+	if (input.segments !== undefined) audit.segments = input.segments;
 	const telemetry = input.telemetry;
 	if (telemetry) {
 		if (telemetry.model !== undefined) audit.model = telemetry.model;
@@ -302,31 +331,23 @@ export function buildPolishAudit(input: {
 }
 
 /**
- * Extra request fields for the polish call, or nothing when the model has no thinking to turn
- * off.
- *
- * Short transcripts keep thinking on: it is cheap there and the wording comes out better.
- * Measured 2026-09-26 on the acceptance corpus, thinking on won exactly the samples this pass
- * exists for — a self-correction merged for +1.71 CER with it on against 0 with it off, and two
- * zh-en term samples +0.08/+0.10 against 0 — while a 161-character transcript spent only 66
- * reasoning tokens in 0.47 s.
+ * Extra request fields for the polish call, or nothing when the model has no thinking to turn off.
  *
- * Long transcripts turn it off: there thinking grows far past the token budget (309 characters
- * needed ~1700 reasoning tokens, 471 characters ~2800-4400, against a budget of 1130-1454), so the
- * answer was truncated and the pass fell back to the raw transcript — intermittently, which is
- * what made a long dictation look unpolished. With thinking off the same input finished in ~1.2 s
- * and spent no reasoning tokens at all.
+ * Thinking is off for every polish call on a reasoning model. It buys quality on short samples
+ * (measured 2026-09-26: a self-correction merged for +1.71 CER with thinking on against 0 with it
+ * off, two zh-en term samples +0.08/+0.10) but it spends a budget nobody can predict: the same
+ * spend that truncated long transcripts also hit short ones - a 76-character dictation burned
+ * ~1,200 reasoning tokens against a 664-1,024 budget and came back truncated after 7.6 s, and
+ * every real-corpus failure recorded on 2026-09-26 was `stop-reason:length` on a transcript of
+ * under 200 characters. On the real corpus the measured correction gain was ~0 either way, so
+ * the pass keeps the wording it can improve and gives up the truncation class entirely.
  *
  * `samplingParams` is applied by OpenAI-compatible adapters only, and the `reasoning` gate keeps
  * the field away from models with no thinking at all.
  */
-export const THINKING_MAX_CHARS = 200;
-
-export function polishSamplingOptions(
-	model: { reasoning?: boolean } | undefined | null,
-	rawLength: number
-): { samplingParams?: { reasoning_effort: string } } {
+export function polishSamplingOptions(model: { reasoning?: boolean } | undefined | null): {
+	samplingParams?: { reasoning_effort: string };
+} {
 	if (!model || model.reasoning !== true) return {};
-	if (Number.isFinite(rawLength) && rawLength <= THINKING_MAX_CHARS) return {};
 	return { samplingParams: { reasoning_effort: "none" } };
 }