feat: add T-I-F reliability helpers (TifScore, evaluate_tif)

[ferhimedamine]

Summary

Part of T-I-F RFC Phase 3 — adds type-safe T-I-F reliability helpers so developers don't have to hand-roll metadata.reliability parsing.

Changes:

• TifScore dataclass in models.py with truth/indeterminacy/falsity proportions, feedback_count, classification property, from_feedback_history() and from_metadata() classmethods
• evaluate_tif(memory_id) convenience method on both DakeraClient and AsyncDakeraClient
• TifScore exported from __init__.py
• 20 unit tests in tests/test_tif.py covering all edge cases and classification thresholds

T-I-F computation:

• upvote/positive → truth
• downvote/negative → falsity
• flag → indeterminacy
• No feedback → {truth=0.0, indeterminacy=1.0, falsity=0.0}

Classification thresholds:

• falsity >= 0.50 → surface_contradiction
• indeterminacy >= 0.50 → ask_clarification
• truth >= 0.70 → confident_reuse
• else → verify_before_use

Related PRs (all 4 SDKs batch)

• dakera-py (this PR)
• dakera-js: feat/tif-reliability-helpers
• dakera-rs: feat/tif-reliability-helpers
• dakera-go: feat/tif-reliability-helpers

🤖 Generated with Claude Code

---

Reviewed-by: Jean-Sébastien Beaulieu (@SeCuReDmE-main-dev) — T-I-F contract parity review

[SeCuReDmE-main-dev]

Phase 3 SDK review note from the RFC side.

This is clean and useful ergonomically: TifScore, evaluate_tif(), metadata parsing, sync/async support, and green CI are all strong.

The one thing I would hold before stabilizing Phase 3 is cross-language parity with the MCP PR and the other SDKs. Right now this Python helper appears to compute raw feedback proportions only, while dakera-mcp#123 adds a thin-evidence base indeterminacy rule when feedback_count < 3. Also, no feedback here is truth=0, indeterminacy=1, falsity=0, which classifies as ask_clarification, while MCP currently returns verify_before_use for no feedback.

I recommend adding shared golden vectors before merge/release, for example:

[]
[upvote]
[upvote, upvote]
[upvote, upvote, upvote]
[downvote, downvote]
[flag, flag]
[upvote x8, downvote x1, flag x1]
[downvote x3, flag x3]

For each vector, Python, JS, Rust, Go, and MCP should return the same truth, indeterminacy, falsity, classification, and feedback_count.

Once that contract is aligned, this PR looks like the right Python-side shape for Phase 3.

[SeCuReDmE-main-dev]

@ferhimedamine final Phase 3 review complete from my side for the Python SDK PR.

I rechecked the current PR state and the DAK-6566 parity fixes. The previous blockers are resolved:

• CI is green.
• PR is mergeable.
• no-feedback now maps to ask_clarification.
• thin-evidence base indeterminacy is present.
• the 8 golden vectors are present and cover the canonical T-I-F v1 contract.
• 3 downvote + 3 flag correctly prioritizes surface_contradiction.
• metadata reliability parsing remains backward compatible with Phase 1 / Phase 2.

This is reviewed from my side. No further requested changes from me.