fix(packaging): nest everything under ks_xlsx_parser + add CI/bench infra

[arnav2]

Why

error: Import "ks_xlsx_parser.pipeline" could not be resolved (reportMissingImports)

Building the v0.2.0 wheel locally and unzipping it confirmed two distinct packaging bugs:

1. pipeline.py and api.py were missing from the wheel — top-level modules under src/, and setuptools.packages.find only picks up packages, so they were silently dropped.
2. Namespace pollution — top_level.txt listed 13 generic siblings of ks_xlsx_parser (models, utils, parsers, …). Installing the package would shadow any other models in site-packages.

CI didn't catch this because the matrix test job uses an editable install — src/ is on sys.path and the wheel-packaging defect is invisible.

What this PR does

• Restructures into a proper nested package — git mv of 13 packages + pipeline.py + api.py + py.typed under src/ks_xlsx_parser/; mechanical rewrite of 47 files' imports; pyproject.toml constrains packages.find to ks_xlsx_parser*; drops PYTHONPATH=src from the Makefile.
• Adds the regression guard CI was missing — scripts/verify_wheel.py builds the wheel, installs in a clean venv, asserts top_level.txt == ['ks_xlsx_parser'] and the public imports resolve. Wired into a new wheel-check CI job and a Verify wheel step in release.yml.
• ks-backend-style CI — split ci.yml into test / wheel-check / lint / typecheck jobs, uv-backed installs. Lint and typecheck are non-blocking with TODO markers (45 pre-existing ruff, 59 mypy — separate cleanup PR).
• Docker benchmark + accuracy tracking — Dockerfile.bench reproduces SpreadsheetBench retrieval scoring. .github/workflows/benchmark.yml runs a 60-instance sample on PRs and the full 912-instance corpus weekly. scripts/append_bench_history.py appends each run to tests/benchmarks/reports/history.jsonl tagged with the git commit.
• Recall failure triage — eval_retrieval.py --emit-failures dumps per-miss diagnostics; scripts/triage_recall.py produces a ranked bucket histogram + exemplar failures so "recall@5 = X" becomes a ranked worklist.
• make install-dev alias (Frank pattern-matched this off ks-backend).

What this PR deliberately does NOT do

• No version bump. Cutting v0.2.1 is a separate step (see docs/RELEASE_PROCESS.md).
• No recall improvements. docs/recall-investigation.md documents the diagnosis framework for the follow-up PRs.
• No fix for the pre-existing ruff/mypy findings — separate cleanup PR.

[arnav2]

Local setup verified end-to-end. Doc added: docs/benchmark-local-setup.md.

Ran python scripts/eval_retrieval.py --parsers ks --sample 60 --emit-failures against the real SpreadsheetBench corpus (downloaded via make corpus-download, 91 MB tarball → 2,726 .xlsx files). Producer side of --emit-failures runs clean. Result:

recall_text@1 = 0.550   recall_text@3 = 0.617   recall_text@5 = 0.633
recall_geometric@5 = 0.283
failure_buckets:
  answer_absent_from_chunks: 11   ← 100% of misses
  present_but_ranked_low:     0
  wrong_sheet:                0
  geometric_no_overlap:       0
  ...all others:              0

Surprising and actionable: 11/11 recall@5 misses fall in answer_absent_from_chunks. Refutes H1 (chunk-size dilution / present_but_ranked_low) and confirms H2 (the parser is dropping cells — extraction gap, not chunking). The next investigation PR should focus on parsers/cell_parser.py + rendering/text_renderer.py, NOT on chunk size tuning.

Concrete example surfaced by triage_recall.py (instance 55260):

Q: "round the amount in the second column to the nearest $5… round the value 367.5…"
answer_position: C1:C2, ground-truth values: ['365', '500']
n_chunks: 1
Top chunk: Sheet2 (1,1):(2,2) — only columns A:B, column C entirely missing from the chunk's range and render_text.

Other observations the doc covers:

• Sample sizes matter — 10-instance run gave recall_text@5 = 0.80 (1 miss); 60-instance run gave 0.633 (11 misses). Decision-grade comparisons need --sample 60 minimum, full corpus for final numbers.
• Mean parse time on the sample: 54 ms (P50 10 ms) — well below the 251 ms full-corpus number in the screenshot, because the sample over-represents tiny instances.

tests/benchmarks/reports/history.jsonl (gitignored) is now populated with both runs locally, demonstrating the row-over-row delta path also works.

The doc covers install → corpus → eval → triage → history → Docker parity → "adding a new bucket" → troubleshooting. Anyone can reproduce this in ~5 minutes after the first corpus + embedding-model download.