… row (Refs #59)

STAGE-B effect-tracking spine. Until now ctx.current_eff was declared
but never read or mutated — effect inference did nothing, and #59's
explicit motivating case ('{Partial} — short-circuit failure, ? on
Result') produced no tracking.

Root-cause, single concrete effect source for tracking-only v1:
- ExprTry with no catch arm (incl. the 'e?' desugar, parser.mly
  expr_postfix QUESTION) unions Partial into ctx.current_eff; a
  handled try (catch arm present) discharges it and adds nothing.
- check_fn_decl now snapshots/resets current_eff around the body,
  then — only when an effect row is *explicitly declared* — requires
  the inferred row ⊆ declared (Effect.eff_subset), erroring with a
  new EffectNotDeclared type_error. Undeclared rows stay permissive
  (tracking-only v1), so blast radius is zero: no stdlib/test uses
  '?' or an explicit row.

Other effect sources (IO calls, Async) remain unconstrained for now —
deliberately scoped; this stands up the accumulate→subsume path they
will plug into, it does not fake full inference.

Tests: test/test_effects.ml — declared /{Partial}+? ok; declared
/{IO}+? rejected (#59 mismatch naming Partial); undeclared permissive;
handled try adds no Partial. Full gate 257/257 (253 + 4, no regression).

Refs #59 (left open for user joint-close per requirements-target rule).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>