claude-config: design per-rule deletion-evidence attribution for unhobble's ledger grammar

[claude]

Context

Signed-off decision Q1 of the context-engineering integration contract
(docs/topics/context-engineering-integration/PLAN.md, 2026-09-01): consequential
standing-instruction deletions need ledger evidence in unhobble's grammar — but the shipped
two-rows-same-cause gate defends RE-ADDS after a full strip, not per-rule deletions during
normal operation. Until an attribution design exists, the consequential deletion tier is
deliberately unclearable (validation record, finding 2).

Proposed work

• Design the per-rule deletion-evidence mechanism: observation window, stumble-to-rule
  attribution rule, same-cause aggregation — mirroring the re-add gate's grammar (rows,
  same-cause, citing commits) in the deletion direction.
• Decide whether it lives as an unhobble extension or a sibling mode; keep one evidence
  grammar for the concern.
• Only then wire audit-instructions' consequential-deletion proposals to require it.

Acceptance criteria

• A written design answers: what counts as a row, over what window, attributed how.
• The two-tier threshold in the integration contract becomes clearable in practice.
• No second evidence grammar is introduced.

References

• docs/topics/context-engineering-integration/PLAN.md (Q1, validation record)
• plugins/claude-config/skills/unhobble/SKILL.md:159-176 (the shipped re-add gate)

---

Generated by Claude Code

[kyle-sexton]

Partly done in #3588 (merged as ae3aa802) — staying open for the wiring

The design this issue asked for now exists on main as
docs/specs/context-engineering-deletion-evidence-attribution.md.
It answers the three questions in the first acceptance criterion — what counts as a row, over what
window, and attributed how — by attributing a stumble to the governed situation rather than to a
named rule, with a co-absence rule for the case where several stripped rules could explain the
same stumble, and it reuses unhobble's existing rows/same-cause grammar rather than minting a
second one. So criteria 1 and 3 are met.

Criterion 2 is not met and this issue should stay open for it. "The two-tier threshold becomes
clearable in practice" needs the last step this issue lists ("only then wire audit-instructions'
consequential-deletion proposals to require it"), and that wiring did not land.

Worth stating precisely, because #3588 did wire audit-instructions and the two are easy to
conflate: what landed is Q2's exception register — a hold on candidates matching a protected
consequence class, which is a categorical carve-out that needs no evidence at all. What is still
missing is Q1's evidence tier: the requirement that a consequential deletion outside those
protected classes cite ledger rows before it may proceed. Different mechanism, different question,
and only the first one is shipped.

The design is no longer memory-tier, so the remaining work no longer depends on this session.

---

Generated by Claude Code