Stop the brain from coercing NA/missing padj to 0 when filtering stats tables

[dannon]

Closes #355

DESeq2 writes NA in the padj column for genes dropped by independent filtering / outlier detection. A bare awk filter like $7+0 < 0.05 coerces those NA rows to 0 ("NA"+0 == 0 in awk), and 0 < 0.05 is true, so every NA-padj gene gets silently counted as significant. In the run that prompted this, one contrast was reported as 8,738 significant genes where the NA-aware count was ~1,560 -- and that wrong number flowed into downstream shared-DEG and TF analyses before anyone caught it. Same class of silent scientific-correctness gap as #318 and #220.

This adds a guidance subsection to the brain's verification-discipline prompt block (buildVerificationDisciplineBlock in extensions/loom/context.ts) -- the shipped system prompt, not the human-facing docs. The rule: when thresholding a p-value/padj/FDR column, drop non-numeric/missing rows explicitly (prefer Python/R with real NA handling, or guard the column in awk) and sanity-check the surviving count against expectation.

A few deliberate choices:

• Prompt note, not a hard guardrail. There's no in-scope way to intercept and parse arbitrary awk/bash; the precedent here (Agent over-claims it compacted the conversation when asked in chat (no real compaction happens) #171, Notebook records a hallucinated analysis date instead of the real run date #268, "Save chat history to a Markdown file" fails with "An unknown error occurred" and writes no file #320) is prompt-text guardrails in context.ts, each with a mechanical test. This follows that.
• Narrowly scoped. It's about coercion of missing values, not "don't use awk" -- legitimate numeric filtering still works, and a column with a deliberate zero-imputation convention is left as the user's call.
• The awk example flags its own limitation. $7 != "NA" only excludes the literal NA; the header row, blanks, and . coerce to 0 the same way, so the text says so and points at Python/R as the reliable default rather than handing over one incantation to cargo-cult.
• Placement: it's in the verification block because the closing count sanity-check is genuine verification and reuses that block's existing row/object-count check. Easy to move to a dedicated scientific-correctness block instead if we'd rather grow that family (Brain should flag basic count invariants in generated analysis summaries (parts must sum to the whole) #318/Brain lacks a cross-assembly gene-ID guardrail — naive locus-tag mapping produced the experiment's only scientific error #220) somewhere of its own.

Ran a skeptical adversarial review pass over the diff; it caught the awk-example-too-narrow and scope-too-broad points above, which are folded in.

Tests: added a case to tests/verification-context.test.ts asserting the load-bearing strings ship in the prompt. Full root suite green (1238 passing), root + app typecheck clean. Not live-eyeballed in a running session.