feat(hooks): tell an unattended session that its kickoff question reaches nobody

[protonspy]

What changed

SessionStart now says one thing when the harness reports the session is unattended: nobody will read a question asked this turn, so take autonomy.md's three kickoff answers as auto / no-wait / the user's own language, record them in frontmatter as assumed rather than answered, and build. The rule gains the matching branch in prose, because two of the three harnesses have no hook surface to deliver it.

Why, and the measurement behind it

design/ponytail-port.md shipped its hook and rule work and skipped its measurement half, leaving its own §Why standing: nothing measures whether the rules earn what they cost. This is that restart at a twentieth of the planned size — two tickets on github.com/dustin/go-humanize pinned at 4d1d908, one model, n=3, 24 headless sessions and 8 blind critics for about $17. design/rule-measurement.md is the writeup, limitations first.

What it found:

• The scaffold costs +14,703 tokens in every request, +42% on the harness's own baseline, with the three scoped rules measurably absent from that figure.
• On a bug fix, blind critics picked the scaffolded arm 3 of 3, twice for the same unprompted reason: it left regression tests the control wrote and then deleted.
• On new work the same scaffold delivered nothing, 3 of 3. It ended every turn holding the kickoff questions and wrote no code, while the control shipped the function every time. Editing the rule's prose changed nothing — 0 of 3 again — because it asks the agent to act on a condition it cannot observe. With this line in front of it: 6 of 6, with the plan, the ticked box, a feat/ branch and a clean scc validate.
• Regression on the small ticket: 7 of 7, diffs of 8–17 lines against the control's 8–20, five of seven under $0.26. The expensive path fires on the ticket that routes to it.

CLAUDE_CODE_SESSION_ATTENDED is undocumented, so exactly one spelling is read and everything else, unset included, is silence. A missed line leaves today's behaviour; a line in front of an attended session would talk somebody out of a question they were right to ask.

The rule pays for its new paragraph with a cut and lands at 54 lines against a cap of 55, so the standing cost does not move.

Verification

• scc validate --checks --pr on the branch: 12 validators, 0 findings; 981 tests, coverage 86.0%.
• go build ./..., golangci-lint run (0 issues), golangci-lint fmt --diff (clean), go test ./... green.

Reviews

Both dispatched on the diff, both returned changes-requested, and all findings are addressed in 85556cf. No blocker or critical.

• scc-code-review — 1 major, 3 minor. The major was the false-positive direction: a session wrongly read as unattended, with no way for a live user to override an already-injected assumption. The minors: the docstring misquoted its own measurement (5 of 5 for 6 of 6), SessionStart's one-line description in scc hooks check had gone stale, and the cut that paid for the new paragraph took with it the reason the CI question in particular belongs at kickoff.
• scc-security-review — 1 medium. An environment variable has no provenance, so a hostile .envrc or devcontainer.json in a clone could tell an attended session nobody is watching and spend the one checkpoint gated exists to provide. It named the unverified link itself, and it was testable: exported as 1 into the shell that launched claude -p, the hook still saw 0 — the harness sets the value rather than inheriting it. One direction of a symmetric mechanism is not a proof of both, so the injected text now ends by saying the conversation wins.

Found, not fixed

One blind critic picked the control on code because the methodology arm had bundled lint-driven rewrites of bigbytes.go, bytes.go and si_test.go into a money-formatting change. The agents recorded the format gate correctly as the check and not the rewrite — the guidance held — but a tree-wide gate on a repository already failing it turns any task into a whole-tree reformat. It happened in 2 of 6 delivering runs and 0 of 3 control runs, which is under the resolution of n=3, so nothing was changed on the strength of it. design/rule-measurement.md §6 records the shape of the fix and what measuring it would cost.

Fourteen of the fifteen rules have still never been told apart from their own absence, and the attended session — the one most people are in — is unmeasured. Both are named in the writeup rather than left implied.