Design docs: the drift PR #58 missed, and a Grid re-walk now that realignment shipped

[JamesKane]

Documentation only — no code changed, so nothing here alters behavior.

Two threads, in two commits.

1. The drift the last audit missed

PR #58 audited all 32 design docs and corrected five headers. A pass the following
day found three more claims it re-affirmed in place without checking. All three run
the same direction as PR #58's findings — the doc claims less than the tree has.

Doc	Said	Actually
RustRewrite_Plan.md §4e	"not yet built"	Built 04728ba, 2026-06-03 — two days after the plan was drafted
project-block-tree.md §11.4 + BACKLOG.md §3.3	"no GUI trigger for the private-Y batch"	The chore table, 331e8cb (PR #47)
project-block-tree.md §11 + BACKLOG.md	project_report debt unpaid	c0994d0, the day before

The §4e case is the interesting one: the section's body already carried the
measured result (13.9 Mb callable chrY short-read vs 1.1 Mb HiFi, the three
private-Y bucket sizes). Only the sub-heading and the "Missing" list contradicted it.
The section disagreed with itself, which is how two consecutive audits read past it.
A status marker that sits away from the evidence is the thing that goes stale.

PR #47 is also the second confirmed instance of a substantial PR shipping and
touching no design doc at all — PR #42 was the first, found by the last audit.

Four docs carried no status header. Three are Scala-era and were missed by PR #58
because they live in documents/, not documents/design/. The fourth,
DecodingUsTreeProvider.md, had no status line of any kind despite being fully
implemented with DecodingUs as the default provider.

2. The Grid re-walk

distributed-compute-grid.md was drafted while realignment was still a plan, and
depended on it. Realignment merged 2026-08-14 — which unblocked the Grid and
overtook one of its four locked decisions. New §11 is the reconnaissance
refresh; D1 and §7.3 are corrected in place, because a locked decision the tree
has already overruled otherwise reads as still binding.

The design's shape survives intact. The cost moved down:

• D1 locked minimap2 via FFI. navigator-align ships minimap2-pure-rs — pure
  Rust, no C toolchain, 99.74 % byte-identical, zero disagreements at MAPQ > 0. §7.3
  split the fleet by OS on that assumption. Windows nodes can take FASTQ units on day
  one; /grid/claim should filter on RAM/disk/threads; the P4 Windows milestone is gone.
• §7.1's "prerequisite refactor" is mostly built — App::analyze_biosample from
  PR Make batch analysis tractable, and fix the Y re-placement bugs it uncovered #47 is already the headless, cancellable, resumable, preflighted per-unit driver.
• mt_terminal is a contradiction, not a gap. The digest demands an exact match;
  analyze_biosample declines to assign mtDNA because it "is not final on CHM13" —
  and the Grid realigns to CHM13. §11.3 gives three ways out and recommends one.
  This is the decision to settle before any P1 code.
• The AppView already resolves ENA at run level and curates it —
  EnaClient::run_files + du-jobs/crawl_project.rs, already splitting BAM/CRAM from
  index files, which is D3's passthrough selector.

Two traps recorded: SV sat in the §7.2 driver at 2–5 h per WGS (vs ~1 h for the whole
rest of the stack) for a signal the digest never compares; and ena.rs cannot mirror
refgenome::download, which sends no Range and hashes SHA-256 where ENA publishes
md5 on multi-GB files.

Net: the critical path is the AppView half — grid.work_unit, the SKIP LOCKED
claim, the validator.

Checks

• scripts/ste-check.py (no arguments, per the known gotcha): 0 files with violations
• No Rust touched, so fmt/clippy/test are unaffected
• Every claim verified by grep against the tree, not by reading a status header

🤖 Generated with Claude Code

https://claude.ai/code/session_01SequryftjcFkLnjQRMKCx3