feat(#510): enforce thesis confidence rules

[adolago]

Summary

• require evidence links for every persisted thesis revision and store the applied confidence assessment on the revision and current thesis record
• add thesis-confidence.v1 to cap conviction from evidence coverage, tool diversity, and linked valuation context while surfacing the rule in thesis snapshots
• document the confidence policy and extend thesis-ledger tests and telemetry for evidence enforcement and conviction downgrades

Testing

• bash scripts/bun-audit-ci.sh
• bash scripts/verify-skills.sh
• bun run --cwd packages/zee typecheck
• cd packages/zee && bun test test/investing/thesis-ledger.test.ts test/investing/research-executor.test.ts
• cd packages/zee && CI=true bun test --reporter=dots --coverage-reporter=lcov
• cd packages/zee && bun run build
• cd packages/zee && ./script/verify-binary.sh
• zee investing thesis status --json

Closes #510

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3f55f8ba94

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve original requested conviction in confidence telemetry

recordInvestingThesisRevision now re-evaluates confidence from input.conviction, but the thesis-refresh automation passes draft.conviction (already capped by the first confidence pass) into this function, so persisted confidence.requestedConviction is flattened to the already-downgraded value. In scenarios where an existing thesis carries high conviction and new evidence only supports medium/low, the stored confidence block and investing.thesis.confidence event no longer record that a downshift occurred, which breaks the new audit trail semantics for requested vs applied conviction.

Useful? React with 👍 / 👎.