protocols/: a reviewed protocol-card library rendered into the tally kit, with the schema and three seed cards (#423, #428, #396)

[mecattaf]

Why

Tom (2026-09-19): aim to always be at issue count zero for dotfiles, and "tool-and-protocol-distill from the exsititign claude runs and turn that into repetitive actins that even halogen can take care of at the atomic level overtime." Design memo (model output, review gate): /home/tom/today/review/2026-09-19/ISSUE-ZERO-DISTILLATION.md sections 4 and 5.

MEASURED on the July to September history: 306 of 890 commits touch one file, 386 stay inside one module, 126 touch only flake.nix or flake.lock. That third is the atomic layer a card can carry.

Outcome

• protocols/README.md describing the card schema: id, class, trigger, inputs, reads, greps, edits (hunk class, anchor, max_lines), may_touch, never_touch, verify (argv, expect_rc, expect_contains), pr (branch, commit_subject, body_template, closes false), budget (prompt_tokens_max, tokens_out, seconds, gpu), tier (origin, validated, current), provenance, history.
• Hunk-class vocabulary, kept small: attr-set-value, attr-list-add, attr-list-remove, line-insert-after-anchor, line-replace-exact, file-add-from-template, table-row-add, pin-bump (the only class allowed to touch flake.lock).
• Three seed cards written by hand from the open issues: protocols/attr-list-add/adapter-allowed-values.json (tally pi adapter: model override not authorized, so a pi job cannot be routed to Halogen or the Qwen plan #423), protocols/attr-list-add/catalogue-refresh.json (Qwen token plan catalogue is stale: endpoint serves 14 models including qwen3.8-flash, the local model's cloud sibling #428), protocols/attr-list-add/failure-surfacing-exclude.json (browser.internal: opening Chrome raises a false xdg-desktop-portal-gtk failure episode #396). Each with a verify argv that is a nix eval or a flake check already declared in flake.nix.
• protocols/templates/pr-body.md in the shape of PRs 389, 413 and 416: changes as bullets, a Validation section naming the checks built, Refs never Closes.
• protocols/issue-classes.tsv: issue number, class (one-shot-done, automate-forever, human-area, live-campaign, keep), tier (fable, opus, cheap, halogen-atomic, hand). Local, not GitHub labels.
• protocols/tools/extract-tool-sequence.py: the deterministic transcript reader (tool_use blocks in order, Bash commands, files edited) used by the distiller.
• home/tally-uplink.nix kit entries card:halogen-atomic (shell runner, cwd the campaign worktree root, env_allowlist HOME PATH LANG TERM, usage_source under the uplink usage directory) and, disabled until Ruling: may the kernel lease a Claude seat for unattended work, and through which row? (rows.md:41-55, tally-b.nix:57-63, D-B6) #362 is answered, card:sonnet-replay and card:opus-origin.
• A flake check that every card parses, has at least one verify entry, names only files that exist, and never lists secrets.nix, modules/secrets.nix or secrets/ in may_touch.

Acceptance

• nix flake check includes the protocols check and it is green on the three seed cards.
• The kit file renders the card entries; claude:headless-style disabled entries are refused by name.
• One card (adapter-allowed-values) has been run once by hand with Halogen on worker-gpu through the shell runner and the branch and receipt are named in a comment here (ruling 22).

Refs #423, #428, #396, #362, #403, tally #61, tally #62.