docs(requirements): add contextual user model requirements with scoring model rationale

Specify the shipped architecture as requirements: dynamic span length
derived from LanguageModel::maxKeyLength() (PR #844), the
ContextualUserModel with absolute-discounting interpolated Kneser-Ney
over bigram contexts (PR #780), and the KeyHandler swap from
UserOverrideModel at the existing observe/suggest call sites (PR #781).

Key specifications:

- Continuation probability normalized per reading:
  P_cont(w|r) = N1+(.,r,w) / sum over w' of N1+(.,r,w'), with
  generalization to unseen contexts gated on >= 2 distinct contexts
  and never force-boosted; minimum suggestion probability 0.25.
- Wall-clock exponential decay count * 2^(-dt/halfLife) with the
  half-life unit in seconds (default 5400 s = 90 min); LRU capacity
  bound (default 500 contexts).
- Two-level scoring with implicit base-LM fallback: insufficient
  evidence yields an empty suggestion and the walk falls back to base
  scores naturally; the model holds no base-LM pointer, eliminating
  the aliasing-shared_ptr lifetime hazard by construction.
- TSV v1 persistence (atomic temp+rename save, validated load,
  serialize() snapshot for off-thread writes) confined to
  contextual-user-model.txt; user phrase files are never touched and
  UserOverrideModel had no persisted data, so no migration is needed.
- The walk algorithm is untouched: the WalkStrategy/fixedSpans
  abstraction and the speculative algorithm variants from PR #779 are
  dropped per maintainer feedback; legacy UserOverrideModel removal is
  a follow-up after #781 bakes.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: tianjianjiang