Build late_perceptual_projection + aggregate; Jon runs on prod; record numbers

Part of #19

Blocked by: #20, #21

## Question

**Build + run.** Port `early_perceptual_projection.py` (+ `_aggregate.py`) into `notebooks/causal46_joined/late_perceptual_projection.py` (+ aggregate) implementing the spec from T1 (per-cell π, late word-end-specific â, no pooling, per-cell null) and T2 (population test, pre-registered criterion). Reuse the substrate: `b4_bootstrap.parquet` for p, `_within_completion.py::bootstrap_endpoint_beta` for â, optionally `b_windows.parquet` for window snapping. Extract shared logic into a helper rather than copy-pasting the early notebook (separate-notebook + extract-shared-logic convention).

Written in the **dev container** (can't execute — `b4_bootstrap.parquet` + epochs are prod-only). Validate as far as the container allows (py_compile, jupytext round-trip, unit checks). Then **Jon runs on prod** (`uv run`, serially); this ticket's answer **records the prod numbers** — the input to the go/no-go (T4). Mirror #16's build→prod-run→record-numbers pattern; include a precise prod run command.

**Deliverable:** the notebook(s) on `late-projection` + prod run instruction. **Resolves** when the population readout is posted here.