Jyotisha/frontend/tests/rectification-grounding-20260927.test.ts

/**
 * TASK-rectification-grounding-20260927 regression tests.
 *
 * T1 / BUG-1055 (recurrence of BUG-588): server fact sentences never take part
 * in the evidence-turn trim, the streamed final state equals the persisted
 * text, and a model sentence with a clock / range / percentage that is not
 * this turn's server fact is dropped whole (P3).
 */
import assert from "node:assert/strict";
import test from "node:test";
import {
  CASE_ID,
  SESSION_ID,
  TURN_ID,
  USER_ID,
  activeFocusFixture,
  candidateSnapshotFixture,
  CANDIDATE_ID,
  SECOND_CANDIDATE_ID,
  conversationSummaryFixture,
  dossierFixture,
  fakeAccounting,
  receiptHandlers,
} from "./rectification-v9-test-support.ts";
import { clientStates } from "./rectification-grounding-support.ts";
import { runV9AgentTurn } from "../src/lib/rectification-agentic/v9/agent-run.ts";
import { buildInferenceState } from "../src/lib/rectification-agentic/core/build-state.ts";
import {
  dropUngroundedFactSentences,
  sentenceNumbersGrounded,
  spokenFactWhitelist,
} from "../src/lib/rectification-agentic/v9/spoken-grounding.ts";
import { RECTIFICATION_USER_COPY } from "../src/lib/rectification-agentic/user-copy.ts";

const BEFORE: [string, string] = ["04:50", "05:10"];
const AFTER: [string, string] = ["04:55", "05:02"];
const RANGE_SENTENCE = "范围从 04:50–05:10 变为 04:55–05:02。";

function inferenceWithRange(range: [string, string]) {
  const state = buildInferenceState({
    range_start: "04:50",
    range_end: "05:10",
    candidates: [
      { id: CANDIDATE_ID, time: "05:02", relative_support: 58 },
      { id: SECOND_CANDIDATE_ID, time: "04:55", relative_support: 42 },
    ],
    events: [],
    probes: [],
  });
  return { ...state, credible_range: range };
}

type TurnCaseOptions = {
  body: string;
  /** Tool results the fake agent reports, in order, after read-case. */
  tools?: Array<{ name: string; result?: unknown; failed?: boolean }>;
  fitPercent?: number;
  /** Flip the stored range to AFTER when this tool result arrives. */
  flipOn?: string;
};

/**
 * Real `runV9AgentTurn` over a fake Mastra agent stream: read-case, then the
 * listed tools, then one model body. The stored snapshot's credible range is
 * BEFORE until the `flipOn` tool result, AFTER from then on (the batch /
 * compare rescore the diagnosis reproduced).
 */
async function runEvidenceTurn(options: TurnCaseOptions) {
  process.env.RECTIFICATION_ALGORITHM_VERSION = "rectification-v5";
  process.env.RECTIFICATION_DECISION_POLICY_VERSION = "rectification-candidate-policy-v2";
  let flipped = false;
  const dossierFor = (range: [string, string]) => dossierFixture({
    latestResult: candidateSnapshotFixture({
      representativeTime: "05:02",
      decisionReceipt: {
        inference_state: inferenceWithRange(range),
        ...(typeof options.fitPercent === "number"
          ? { event_fit_rate: { matched: 4, total: 5, percent: options.fitPercent, band: "high" } }
          : {}),
      },
    }),
    conversationSummary: conversationSummaryFixture({
      activeFocus: activeFocusFixture({
        askedTurnId: null,
        intent: "collect_method_evidence",
        expectedAnswerSchema: { prompt: "你大概是哪一年搬的家？", collect: true },
      }),
    }),
  });
  const accounting = fakeAccounting({
    ...receiptHandlers,
    get_agentic_rectification_case_dossier: () => dossierFor(flipped ? AFTER : BEFORE),
    append_agentic_rectification_turn: () => ({ turn_id: TURN_ID }),
    finalize_agentic_rectification_turn: () => ({ turn_id: TURN_ID, status: "completed", idempotent: false }),
  }, { fallback: () => null });
  const flipOn = options.flipOn ?? "rectification-compare-candidates";
  const tools = options.tools ?? [{ name: "rectification-compare-candidates" }];
  const chunks: Array<{ type: string; payload?: Record<string, unknown> }> = [
    { type: "start" },
    { type: "tool-call", payload: { toolName: "rectification-read-case", args: { caseId: CASE_ID } } },
    { type: "tool-result", payload: { toolName: "rectification-read-case" } },
  ];
  for (const tool of tools) {
    chunks.push({ type: "tool-call", payload: { toolName: tool.name, args: { caseId: CASE_ID } } });
    chunks.push(tool.failed
      ? { type: "tool-error", payload: { toolName: tool.name, error: new Error("engine_request_failed") } }
      : { type: "tool-result", payload: { toolName: tool.name, result: tool.result ?? {} } });
  }
  chunks.push({ type: "text-delta", payload: { text: options.body } });
  chunks.push({ type: "finish" });
  const agent = {
    stream: async () => ({
      fullStream: (async function* () {
        for (const item of chunks) {
          if (item.type === "tool-result" && item.payload?.toolName === flipOn) flipped = true;
          yield item;
        }
      })(),
      totalUsage: Promise.resolve({ inputTokens: 10, outputTokens: 20 }),
    }),
    getSkill: async () => ({ name: "jyotish-birth-time-rectification", instructions: "skill" }),
  };
  const emitted: Array<Record<string, unknown>> = [];
  const result = await runV9AgentTurn({
    userId: USER_ID,
    caseId: CASE_ID,
    sessionId: SESSION_ID,
    requestId: "aaaaaaaa-bbbb-4ccc-8ddd-eeeeeeeeeee1",
    action: "evidence",
    message: "2016年9月离开家去北京工作",
    modelName: "m",
    accounting: accounting.client as never,
    billing: { reserve: async () => ({ success: true, status: 200 }), complete: async () => true, release: async () => true },
    emit: (event) => { emitted.push(event as never); },
    buildAgent: async () => agent as never,
  });
  const finalize = accounting.calls.filter((call) => call.fn === "finalize_agentic_rectification_turn").at(-1);
  const persisted = String(finalize?.args.p_assistant_message ?? "");
  return { result, emitted, persisted, states: clientStates(emitted) };
}

/** A fact sentence, once shown, must stay in every later client state. */
function assertNeverWithdrawn(states: readonly string[], fact: string) {
  const first = states.findIndex((state) => state.includes(fact));
  if (first < 0) return;
  for (const state of states.slice(first)) {
    assert.ok(state.includes(fact), `fact withdrawn after being shown: ${JSON.stringify(states)}`);
  }
}

test("BUG-1055 repro 1: two-sentence body keeps the server range sentence, stream ends where persist does", async () => {
  const { result, persisted, states } = await runEvidenceTurn({
    body: "记下了：2016 年 9 月去北京工作。这件事我拿去和星盘对照了。",
  });
  assert.equal(result.ok, true);
  // Before: the trim kept only the first sentence and the range sentence was cut.
  assert.equal(persisted, `记下了：2016 年 9 月去北京工作。这件事我拿去和星盘对照了。${RANGE_SENTENCE}`);
  assert.equal(result.answerText, persisted);
  assert.equal(states.at(-1), persisted);
  assertNeverWithdrawn(states, RANGE_SENTENCE);
});

test("BUG-1055 repro 1b: three-sentence body is trimmed to one, range sentence still whole", async () => {
  const { persisted, states } = await runEvidenceTurn({
    body: "记下了：2016 年 9 月去北京工作。这件事我拿去和星盘对照了。接下来再对一件事。",
  });
  assert.equal(persisted, `记下了：2016 年 9 月去北京工作。${RANGE_SENTENCE}`);
  assert.equal(states.at(-1), persisted);
  assertNeverWithdrawn(states, RANGE_SENTENCE);
});

test("BUG-1055 repro 2: a first sentence carrying the stale range is dropped whole, not persisted", async () => {
  const { persisted, states } = await runEvidenceTurn({
    body: "记下了：2016 年 9 月去北京工作，目前范围在 04:50–05:10。接下来再对一件事。",
  });
  // Before: 「记下了：2016 年 9 月去北京工作，目前范围在 04:50–05:10。」was persisted.
  assert.doesNotMatch(persisted, /目前范围在 04:50–05:10/);
  assert.equal(persisted, `接下来再对一件事。${RANGE_SENTENCE}`);
  assert.equal(states.at(-1), persisted);
  assertNeverWithdrawn(states, RANGE_SENTENCE);
});

test("BUG-1055: an invented fit rate sentence is dropped; a server fit rate sentence is kept", async () => {
  const invented = await runEvidenceTurn({
    body: "记下了：2016 年 9 月去北京工作。目前吻合率 90%。",
    fitPercent: 80,
  });
  assert.equal(invented.persisted, `记下了：2016 年 9 月去北京工作。${RANGE_SENTENCE}`);
  const grounded = await runEvidenceTurn({
    body: "记下了：2016 年 9 月去北京工作。目前吻合率 80%。",
    fitPercent: 80,
  });
  assert.equal(grounded.persisted, `记下了：2016 年 9 月去北京工作。目前吻合率 80%。${RANGE_SENTENCE}`);
});

test("BUG-1055: the post-rescore range and representative minute may be repeated; only mismatches go", async () => {
  const kept = await runEvidenceTurn({
    body: "记下了：2016 年 9 月去北京工作。代表分钟是 05:02，范围 04:55–05:02。",
  });
  assert.equal(kept.persisted, `记下了：2016 年 9 月去北京工作。代表分钟是 05:02，范围 04:55–05:02。${RANGE_SENTENCE}`);
});

test("BUG-1055: when the whitelist leaves no model body the batch recap stands in", async () => {
  const { persisted } = await runEvidenceTurn({
    body: "目前范围 04:50–05:10，吻合率 90%。",
    tools: [{
      name: "rectification-record-evidence-batch",
      result: {
        items: [{ index: 0, outcome: "accepted" }],
        accepted_recaps: [{ display_date_label: "2016 年 9 月", event_phrase: "去北京工作" }],
        accepted_count: 1,
        rescore: { status: "completed" },
      },
    }],
    flipOn: "rectification-record-evidence-batch",
  });
  assert.equal(persisted, `记下了：2016 年 9 月 去北京工作。${RANGE_SENTENCE}`);
});

test("BUG-1055: 这次没有重新比较 survives a multi-sentence body and replaces the range sentence", async () => {
  const { persisted, states } = await runEvidenceTurn({
    body: "记下了：2016 年 9 月去北京工作。这件事拿去对照了。还有一句。",
    tools: [{
      name: "rectification-record-evidence-batch",
      result: {
        items: [{ index: 0, outcome: "accepted" }],
        accepted_recaps: [{ display_date_label: "2016 年 9 月", event_phrase: "去北京工作" }],
        accepted_count: 1,
        rescore: { status: "failed", error_code: "engine_request_failed" },
      },
    }],
    flipOn: "never",
  });
  assert.equal(persisted, `记下了：2016 年 9 月去北京工作。${RECTIFICATION_USER_COPY.rescoreSkipped}`);
  assert.equal(states.at(-1), persisted);
  assertNeverWithdrawn(states, RECTIFICATION_USER_COPY.rescoreSkipped);
});

test("BUG-1055: 候选比较这次没跑成 survives a multi-sentence body", async () => {
  const { persisted, states } = await runEvidenceTurn({
    body: "2016 年 9 月去北京工作这件事拿去对照了。它落在那年的变动里。还有一句。",
    tools: [{ name: "rectification-compare-candidates", failed: true }],
    flipOn: "never",
  });
  assert.equal(persisted, `2016 年 9 月去北京工作这件事拿去对照了。\n\n${RECTIFICATION_USER_COPY.compareFailedRetry}`);
  assert.equal(states.at(-1), persisted);
  assertNeverWithdrawn(states, RECTIFICATION_USER_COPY.compareFailedRetry);
});

test("P3 whitelist is a gate on whole sentences: clocks, ranges and percentages", () => {
  const whitelist = spokenFactWhitelist({
    credibleRange: ["04:55", "05:02"],
    representativeTime: "04:58",
    fitPercent: 79.6,
  });
  assert.deepEqual(whitelist.percents, [80]);
  assert.equal(sentenceNumbersGrounded("范围 04:55–05:02。", whitelist), true);
  assert.equal(sentenceNumbersGrounded("范围 04:55 到 05:02。", whitelist), true);
  assert.equal(sentenceNumbersGrounded("代表分钟 04:58。", whitelist), true);
  assert.equal(sentenceNumbersGrounded("吻合率 80%。", whitelist), true);
  assert.equal(sentenceNumbersGrounded("吻合率百分之80。", whitelist), true);
  assert.equal(sentenceNumbersGrounded("2016 年 9 月入学。", whitelist), true);
  assert.equal(sentenceNumbersGrounded("范围 04:50–05:02。", whitelist), false);
  assert.equal(sentenceNumbersGrounded("范围 04:55–05:10。", whitelist), false);
  assert.equal(sentenceNumbersGrounded("代表分钟 05:00。", whitelist), false);
  assert.equal(sentenceNumbersGrounded("吻合率 90%。", whitelist), false);
  assert.equal(sentenceNumbersGrounded("吻合率 80.5%。", whitelist), false);
  // Whole sentence out; the neighbours stay untouched (no in-sentence deletion).
  assert.equal(
    dropUngroundedFactSentences("记下了：2016 年入学。范围在 09:00–09:05，吻合率 90%。还有吗", whitelist),
    "记下了：2016 年入学。还有吗",
  );
  assert.equal(dropUngroundedFactSentences("范围在 09:00–09:05。", spokenFactWhitelist({})), "");
});