[mspinola]

Closes out docs/handoffs/2026-08-02-coverage-reporting.md. The body is preserved verbatim; only the Status: line is replaced, per the append-only rule.

1. The status line was missing a correction

It read complete (PR #15), corrected same day (PR #17). There was a third, PR #19 (7ff5d06), and it is the only one of the three that changed what the module says about itself: the docstring table still claimed 058643 "dies at the price join", when drops_at returns extremity_z and price is 37 of 880 rather than 0.

It also settled a contradiction a reader would otherwise hit head-on. §B18 says the two lumber codes "fail at the SAME rung" — true of the terminal rung, false of drops_at, which is two rungs earlier for 058643.

2. §B30: the rename trap has a mirror image

coverage keys on market_code, which kills the four phantoms (one code, several names). It does not handle the opposite shape, and nothing had checked:

	shape	handled by
phantom	one code, several names	coverage, by keying on the code
split	one instrument, several codes	continuity, merge_migrated_codes, not coverage

§B26 and §B27 established after this handoff closed that the two unscoreable markets are the two halves of one migrated contract: 058643 and 058644 overlap in 7 weeks of 1051 and both carry contract symbol LBR. Nobody re-ran the coverage report afterwards.

So is "2 of 27 score nothing" real, or an artifact of the split? The suggestive number says artifact: separately the codes hold 37 and 178 priced weeks, but merged they hold 208 contiguous priced weeks, twice the 104-week extremity window.

Measured end to end, it is real:

rung	058643	058644	merged
price	37	178	208
extremity_z	0	75	96
illiquidity	0	75	92
crowding	0	0	0
damage_sell non-null	0	0	0

Every rung rises substantially and the verdict does not move. C = pct(z) stacks a second three-year window on the 96 z values and 96 does not fill it. Lumber is unscoreable because it has four years of prices against a measure needing six, not because its code changed. The headline moves from "2 of 27" to "1 of 26" purely by counting one instrument once.

Recorded because the wrong answer survives one round of checking: merging only the per-category half also returns crowding=0, for the unrelated reason that build() still reads the unmerged store and every per-market rung reports 058644 alone. Two routes to the same number, one of them meaningless.

This also vindicates §B18 on the point it was least sure of: coverage and the PCA disagreeing about lumber "is correct rather than a contradiction". That was an argument; this is the measurement behind it.

No code change follows. Teaching coverage to merge migrated codes would change the report's row count and no conclusion in it, and this handoff's own Scope puts "changing any window, minimum or threshold" out of bounds. It is a build item for whoever wants it, now backed by a measurement of what it would and would not buy.

3. Fixed a breakage I caused

Merging cotdata#93 deleted the function reproduce.py's flow_equivalence block calls, so §B29's figures had a reproducer that raised AttributeError. Found by running the full reproducer rather than only the new block.

It now detects the removal and prints the figures as historical with the reason, matching what test_flow_equivalence.py already does. A reproducer that crashes once the change it recommended lands reads as a broken measurement rather than a completed one.

Both flagged decisions stand

The two decisions the handoff refused to default are unchanged four PRs later: a zero-scoring market still stays in the panel, and LADDER is still a hand-written list guarded by test_the_ladder_covers_every_column_the_composite_consumes. That guard has earned its place since: §B18 found the ladder skipping three rungs, and the test now checks the columns add_composite computes, not only those it emits.

Checks

• 433 tests pass, ruff clean
• Prior reproducer output byte-identical; the new lumber_is_one_instrument block is additive
• Handoff body preserved verbatim, only Status: replaced
• No em dashes in added prose; every quoted figure has a reproducer

🤖 Generated with Claude Code