feat(query): opt-in --recency flag to down-weight stale facts

[TPAteeq]

What

Adds an opt-in --recency flag to graphify query (and the equivalent
recency / half_life_days fields on the MCP query_graph tool) that
multiplies each matched node's search score by a time-decay factor, so newer
facts rank ahead of otherwise-equal stale ones. Default query output is
byte-for-byte unchanged when the flag is off.

This is the small, low-risk slice of #1650. The larger, riskier piece — fact
supersession / temporal-validity invalidation (valid_from/valid_to,
marking an old fact as superseded by a newer one) — is deferred to a
follow-up and is intentionally not implemented here.

How

• Reuses the existing half-life math. graphify reflect already has a pure
  _decay() (halves every _DEFAULT_HALF_LIFE_DAYS = 30); serve.py now
  imports those two pure names and decays query scores on the same curve. No
  coupling to the reflect learning sidecar.
• Recency signal precedence (_node_recency_weight): captured_at (ISO
  datetime, present only on ingested docs) → else the source_file's on-disk
  mtime resolved under the repo root → else 1.0 (neutral). Code/AST nodes
  carry neither, so recency is a no-op for them.
• Threading: _query_graph_text(..., recency, half_life_days, now, source_root) → _score_nodes(...). The decay multiplier is applied only
  inside if recency:; when the flag is off, every code path is identical to
  the pre-Feature: temporal validity on facts + recency weighting in query (living corpora return stale facts at full weight) #1650 scorer (verified by byte-identical tests).
• Determinism / testability: a query is inherently "now"-relative, which is
  exactly why this stays behind an opt-in flag. _score_nodes /
  _query_graph_text accept an optional explicit now anchor so tests inject
  ages instead of depending on the wall clock.
• _pick_seeds' per-term coverage guarantee is left age-neutral on purpose: a
  single old match for a query term shouldn't be starved out just for being old.

CLI / MCP surface

graphify query "<question>" [--dfs] [--context C] [--budget N] \
    [--recency] [--half-life-days N] [--graph path]

MCP query_graph gains recency (boolean, default false) and
half_life_days (number, default 30, only used when recency=true).

Tests

tests/test_serve.py and tests/test_query_cli.py:

• flag off → _score_nodes / _query_graph_text output byte-identical to
  today, even when captured_at is present;
• flag on → the newer of two equally-matching nodes is promoted (ranking
  and the Start: seed order shift toward newer);
• captured_at precedence, on-disk mtime fallback, neutral weight when neither
  is present, future-date clamp, and _source_root_for path derivation;
• CLI --recency flip and --half-life-days parsing/validation.

Ages are injected via captured_at + an explicit now anchor (unit tests) or
decades-apart dates (CLI tests, which have no now-injection), so nothing is
wall-clock dependent.

Run:

uv run pytest tests/test_serve.py tests/test_query_cli.py \
    tests/test_serve_http.py tests/test_reflect.py tests/test_querylog.py -q

Result: 165 passed, 1 skipped.

Deferred (follow-up, not in this PR)

• Fact supersession / temporal-validity invalidation (valid_from/valid_to,
  superseded-by links). Only the query-time recency weighting is implemented here.

Refs #1650