Match math.sum valid-value windows and canonicalize validation reports

[luisleo526]

What changed

• make math.sum ignore na, remain na until length valid observations
  exist, and then retain the last-length valid-value window
• preserve same-bar recompute and rollback semantics with focused coverage
• expose the canonical verifier as structured analysis and make validation
  reports consume that single scoring path
• make run_corpus.sh regenerate the report and run its drift guard
• advance the corpus submodule to merged corpus PR fix(security): LTF script_tf upper bound (req must be finer than script) #9

Why

The runtime previously returned finite partial sums during warmup. On a
chart-window restart this shifted KAMA state and downstream crossings. The
correct valid-value window, composed with the independently captured chart
origin in the corpus, reproduces the immutable 4,979-trade oracle exactly.

Closeout also found that the Markdown report duplicated an older tier scorer,
so it could disagree with the canonical verifier, and run_corpus.sh did not
actually regenerate the report despite documenting that behavior. This change
removes the duplicate rubric and restores the report step to the main corpus
workflow.

Impact

• Kanuck KAMA recurrence: Strong to Excellent, 4,979/4,979 exact
• no comparison tolerance, tier threshold, TV tape, Pine source, or global
  chart-origin rule changes
• old-versus-refactored canonical verifier: zero tier or verbose-output changes
  across all 252 probes

Validation

• CTest: 100/100
• C ABI runtime check: pass
• codegen: 1,537 passed / 2 skipped
• full clean-upstream corpus: 252/252 at 241 Excellent / 10 Strong / 1 expected
  Anomaly, with no downgrade
• narrowed committed corpus verifier: 252/252 at 240/11/1
• report/verifier lock-step: 252/252
• forced no-cache 412-strategy collateral sweep: all rows accounted, 0 DOWN
• fresh-context R7: PASS