evidence for the rows an answer is not sure of

[swapnilpaliwal-sd]

Rows an answer is not sure of ([by name], [one of a set], [registered], [text]) come without the code that decides them, so readers re-open the file for about half of all answers, and wrong name-matches get accepted.

• New shared layer scripts/ax_evidence.py, used by impact, path, context, changed and test-impact in grep, prose and --json output, by the MCP tools and by the post-edit hook note.
• For each of the 5 strongest non-exact rows it adds the call line and the deciding line: the parameter or field that types the receiver, the DI binding, the dispatch key, or the registration constant. Each row gets at most 2 lines of at most 160 chars. The remaining rows are only counted, and exact rows get nothing.
• only_through: how many callables and tests are reached only through that row. --drop <file:line> and --exact ask again without it (MCP drop / exact).
• alongside rows collapse to a count when evidence is on (--alongside lists them).
• Switch for an A/B test: AXIOMCODE_EVIDENCE=on|off, --evidence/--no-evidence, MCP evidence. Default off, so every answer is unchanged unless it is turned on.
• JSON rows gain evidence: {call, decider} and only_through: {callables, tests}.
• Golden case in five languages (TS name-match on a Map field / an any parameter, dispatch key, registration constant; JS; Python __init__ param; Java Function.apply; C# dynamic field). Controls: a resolved-only answer is byte-identical with the switch on and off, and switch off is byte-identical to base.

Checked: golden case 14/14 (0/14 on base); language suites ts 98/0, js, py 42/0, java 79/0, cs pass; CLI cases fail the same 9 checks as base; TS probes 31/74 before and after, 0 regressed; product edges 9977 = 9977; p90 answer growth with evidence on: 0–1.42 KB across the five products (limit 1.5 KB).