reconcile: the card after a second read (ACE-128) - #310

Spec: ACE-128 (agami-sdlc main, projects/agami-core-extension/specs/reconcile-evidence/ACE-128-card-second-read.md)

Do not merge without manual review. Stacked on #309 (ACE-127); merge bottom-up: #288 → … → #307 → #309 → this.

Why

Sandeep read the result-and-fix card on a real run and found seven things. Each is fixed here with a test.

Second commit. After Sandeep: "we dont need this explanation with the diff ... use github type formatting if needed" and "the formatting of the tool on a windows machine is not great". The columns row carries no sentence: a column only yours has reads as removed (red, a minus), one only agami's as added (green, a plus), a pair with two names is marked in place with the partner's name on hover (renamed on the diff row). Fonts: the pages named ui-monospace, monospace, which Windows browsers resolve to Courier New, and a Mac-first body stack; every monospace cell now names Cascadia Mono, Cascadia Code and Consolas before the generic, the body names Segoe UI first, the marks and arrows draw from Segoe UI Symbol where it exists, and the value cells sit at 13px.

Third commit. After Sandeep: "the checks panel should be collapsed by default" and "make sure that the colors are consistent. follow a single palette". The checks panel starts closed on every card. One palette, stated once at the top of the shared stylesheet and used by every pill, mark, chip and diff token: green passed/match/keep; red your query; violet the semantic model; amber the question or a check that could not run; blue agami's side (new token --agami: a column only agami returned, add an example, ask again, the suggested decision); gray noted. A legend under the filter rows says the six meanings.

Fourth commit. Sandeep found the mirror cost of the rename fix: identical column names read as a column difference when the row counts differ (13 vs 759), because values cannot pair across vectors of different length. The comparator now reports no pairs and no extras when the counts differ, and the items verb compares columns by data only when a values comparison ran (same counts); otherwise it falls back to names, so identical names are one column set and the rows check alone carries the difference. Tests on both sides.

Review round (/agami-sdlc:review over #304 to #310, three passes). Fixed in the fifth commit: a "same answer, different query" row was never offered for keep; a date window the ledger had confirmed read "could not check" and counted twice; the relative-window fold had no depth budget and could raise out of read_claims on a pathological statement (now eight steps, then None). Nits taken: resume refuses a corrupt checkpoint; a definitional claim that could not be read makes the query fact "not comparable"; a match with an extra column of agami's stays a match and the sentence names the column; a claim that differs is its own amber state (≠) so red keeps one meaning; the intake parser refuses a non-intake rows file and ties the block to the run --out names; intake refuses a file over 20 MB or a CSV cell over the field limit with one line; the page shows only the first line of an error; example is never offered on a row whose ledger holds a mistake; fix words live in one place; the Node page test asserts on CI; em-dashes out of three titles; a tautological assertion and a duplicated check removed. Specs trimmed to decisions in force (ACE-126's build inventories moved here; ACE-128's superseded lines merged; ACE-125 names resume). Declined: deleting the four-beat fallback for older items files (ACE-123's own tests still use that shape; a clean-up when #303 lands); a .gitignore line for a staging folder that should never be created inside a worktree again; --run required on the intake parser (it now derives from --out). Escalated to Sandeep: the rubric pass judges #307 too large (1757 lines, 3.5x the guideline) and asks to lift its library changes (the claims reader's relative windows, the receipt's aggregate flag, the ledger's metric rule) into their own PR between #305 and #307.

Sixth commit. After Sandeep: "the numbers on the Fix pills should get updated, when a filter on Result." The chip counts are facets: each row counts the rows the other filters and the search leave, and a chip whose count fell to zero stays, dimmed, so it can still be toggled.

Decisions

• Columns are compared by the values they carry, never by name. The comparator already paired them that way; the page now reads the pairing.
• example is a decision of its own. "Change the semantic model" was the wrong door for a prompt example.
• Prefill is a start, not a decision. A box the person has typed in is never overwritten.

Tests

• Comparator: renamed and extra columns as pairs and extras; scalar pair; other levels report nothing. Claims: bare and qualified spellings agree on one table; two tables keep a bare column bare. Items: columns by data with pairs, a fully renamed table as a match, one source for change text and prefill. Parser: the example decision with and without words. Page: question-first head, folding checks, the example option, prefill; the Node smoke test still loads every page. 481 library tests and the reconcile suites green.
• Full uv run dev.py check on the final commit (7d14fa3): ruff clean, vendored library in sync, 5888 passed, 12 skipped, 1 failed, the failure being tests/test_golden_run.py::test_a_client_that_cannot_be_found_still_fails_as_a_generation, the known local failure on this machine.

History note (review round). Replayed on the rewritten #309 (its staging folder removed); content unchanged.

Rebase note (2026-09-13). Replayed onto main at 99e44a8 with every branch below it; content unchanged, force-pushed with lease.

🤖 Generated with Claude Code