feat: implement from a typed issue — --issue=<id>, issue_show/issue_verify (#173)

[alpibrupa]

Part of #173 (items 1–3; item 4, the agent-refined acceptance flow with human approval, is a follow-up). Consumer side of lex-lang epic alpibrusl/lex-lang#949.

What

• --issue=<id> ["guidance"]: fetches the issue (lex issue show) and renders its acceptance as the task: exact signatures to ADD/CHANGE/REMOVE, the examples as the oracle, or the failing example to fix (src/issue_contract.lex, pure, covered by examples). Whatever the model claims, the run ends with lex issue verify and prints [ISSUE_VERDICT]\t<verified|failed|inconclusive|unavailable>\t<id> as its last line.
• Provenance: .lex/intent/issue = <session>\t<issue> binds the session to the issue, so every clean .lex write publishes with --intent-issue. The chain is issue → intent → ops → attestation. The file is honoured only for the session that wrote it, so a leftover file can't put its issue on a later session's ops.
• Tools issue_show / issue_verify in the build and minimal toolsets, so the fix loop iterates against the issue's own oracle. A failed verdict is an answer (ok:true, exit 1) with the detail naming the wrong signature or example. It is not a tool error.
• Smoke gate: added unknown command to the refusal vocabulary. A toolchain older than the command a tool calls prints exactly that, and none of the existing phrases matched it.
• Toolchain 0.11.32 → 0.11.66 (lex issue needs ≥ 0.11.50). I didn't use 0.11.67: lex-lang #1005 runs git ls-remote once per import, with no memoization. That makes lex check src/tools/index.lex take about 2 minutes (8s CPU), and CI checks ~115 files one by one. I measured it with a logging git shim: 21 ls-remotes for one small file, lex-schema alone 13×.

Verification

• Local replica of every CI step on 0.11.66: type-check of all files, fmt, doc-sync, smoke (SMOKE OK — 29 tools, including issue_show/issue_verify), tests 4/4, bar gate, effect minimality, and the manifesto twin. All green.
• Negative control: a deliberately wrong examples {} expectation in issue_contract.lex is rejected by lex check (example_mismatch), so the new examples really run.
• Live end-to-end with local qwen3.8:27b-mlx (OLLAMA_THINK=false) on a fresh package:
  • Set up a typed_delta issue (digit_sum(n :: Int) -> Int, 3 examples) that fails at head.
  • lex-code --ollama --issue=<id> wrote digit_sum with the issue's examples, called issue_verify itself, and ended with [ISSUE_VERDICT] verified.
  • An independent lex issue verify also returned verified.
  • Both new ops carry the intent whose issue_id is the issue. The baseline op's intent has none.

🤖 Generated with Claude Code