Release v1.2.1 — Scorer Hardening & Rust Safety Lints · anulum/director-ai

What's New

strict_mode parameter — disables heuristic fallbacks when NLI is unavailable.
Returns neutral 0.5 instead of keyword heuristics, preventing misleading scores
in production deployments without torch installed.
```
scorer = CoherenceScorer(strict_mode=True, threshold=0.6)
```
Configurable scoring weights — w_logic and w_fact constructor params
(default 0.6/0.4) let you tune the logical vs. factual divergence balance
for your domain.
```
scorer = CoherenceScorer(w_logic=0.3, w_fact=0.7)  # fact-heavy for RAG apps
```

#![deny(unsafe_code)] enforced on 5 of 6 Backfire Kernel crates
(backfire-types, backfire-core, backfire-physics, backfire-ssgf,
backfire-observers). The compiler now rejects any future unsafe blocks.
backfire-ffi documents why deny(unsafe_code) can't apply (PyO3 proc
macros generate unsafe internally) and carries FFI safety invariants.
4 safety invariants documented on backfire-core: halt irreversibility,
NaN/Inf lethality, independent halt mechanisms, zero hot-path allocations.

"Known Limitations" section in README — honest about heuristic accuracy,
summarisation weakness, single-document NLI, and weight sensitivity.
Full CHANGELOG entries for both v1.2.0 and v1.2.1.

4 new tests: strict_mode logical/factual neutral return, heuristic mode
divergence, custom weight assignment.
378 tests passing (8 skipped — optional deps).