Add implementation plan: review + phased build plan

[Keeeeeeeks]

Summary

Adds pipeline-spec/05-implementation-plan.md — turns the v5 spec into a runnable, cluster-scale plan.

• Adversarial review of the five pipeline-spec docs, verified against the actual v4 code (false-match test, is_cached resume, create_repo-only upload, --gpus=1 ghx4 for a CPU job).
• Gap analysis vs PI feedback (false-match, consistency, CPU/GPU, DONE markers, dataset availability).
• Data-access matrix + HF/MMU inventory aligned to the UniverseTBD "Multimodal Universe HATS" collection; corrected for org namespaces (UniverseTBD/ vs hugging-science/ vs LSDB/), with APOGEE DR17 and Legacy-south now on HF, and a flag to verify mmu_gaia_gaia full-vs-preview.
• Reuses Smith42 pre-cross-matched sets where they cover a pairing, then lsdb's built-in crossmatch for HATS catalogs; convert non-HATS sources to HATS with hats-import and flag any we can't (sec 4.5).
• Locked decisions: galaxies-first / quality-gated volume; data-gen on x86 Delta cpu + analysis on DeltaAI (shared /taiga); Parquet+fixed_size_list.
• Fixes: DeltaAI no-CPU-partition (C1), apply_space_motion distance trap (H1), 2-D work model + finalize barrier (H2), manifest build-or-attach (H3), Lustre small-file storm (H4).
• Phased P0->P2 plan with modules, scripts, and per-phase acceptance criteria; open decisions for PI sign-off.

Notes

• Stacked on Add pipeline-spec: v5 cross-match dataset methodology + script spec #1 (add-pipeline-spec) so the diff shows only the new doc. Retarget to main after Add pipeline-spec: v5 cross-match dataset methodology + script spec #1 merges (gh pr edit 2 --base main).
• External facts verified via research (HF collection/pages, lsdb/hats API, NCSA Delta/DeltaAI docs, Legacy DR10 docs).

Test plan (Phase 0, before the expensive full run)

• PI reviews the three open decisions: galaxy-image source (HSC-only vs +Legacy), volume ambition, match strictness.
• Confirm the compute nodes (not just the login node) can reach the internet (HuggingFace / S3). If not, switch to pre-staging data via Globus.
• On one small patch of sky, confirm our galaxy image-to-spectrum matches agree with Smith42's already-matched catalog for that patch (proves the matching is correct before committing cluster hours).