Hypothesis v3 quality, readable Korean text, retry on by default, setup and trace text fixes

[hyperdrivebeep]

When a trace row fails, THOTH now proposes hypotheses whose expected-result table can actually tell them apart, keeps Korean answers in Hangul, and no longer leaves a workspace stuck when the web consent was never chosen.

What changes

• Hypothesis contract v3 (opt-in per project). Each hypothesis gets a prediction on every test. Each column assumes its own hypothesis is the cause. Tests are chosen to fix or measure conditions a hypothesis could interact with, and the hypotheses in a portfolio exclude one another. A separate narrow call fills cells the generator left open and records what it answered, including a one-sentence reason for each remaining "모름".
• Readable Korean. Every model role is told not to use Chinese characters or kana absent from the question or quoted evidence. For v3 hypotheses, a readable list item in the wrong script is dropped with a record, unless it is a quotation or the last item a record needs. Single texts are only flagged. The katakana middle dot is shown as the Korean one.
• Interrupted model calls retry by default. The existing retry (at most two extra sends, 2 s then 4 s, shared with the transient-429 retry) is now on for projects without a saved choice. Broken streams record remote request identifiers.
• Trace investigations get comparison conditions and CSV rows. Trace and CSV import messages choose 이/가, 을/를, 은/는 by sound instead of showing "이(가)".
• Settings. A local workspace that already has projects can choose the web consent in Settings. The composer says which setup step blocks a new investigation. Account status says "execution not yet verified" once.
• Demo script. It reads field-specific names from the demo manifest. The radar demo output is byte-identical, and a test pins it.

Validation

Clean checkout of this branch, same tool versions as CI:

• Python full suite: 3,088 tests, 3,086 passed, 0 failed, 2 skipped (both skips state their own reason).
• Web: 718 tests passed, 1 skipped; lint, typecheck and build pass.
• ruff, architecture, module budget, public protocol parity, atomicity coverage and source manifest checks pass. README quick start serves the first page.
• Live model runs on synthetic data: v3 met its fixed criteria in 3 of 3 runs on the radar demo and 3 of 3 runs on a second synthetic dataset from a different field. One end-to-end UI run from a trace row also met them.

Known: Claude/xAI login-path tests emit UnicodeDecodeError thread warnings from subprocess output on a Korean-locale Windows host. These are warnings, not failures, and are unrelated to this change.