TASK-SKILL-117: architectural-spike-author / -audit

§1 - Description

A spike is the time-boxed investigation that turns "we have 2+ plausible architectures" into evidence an ADR can cite. The ADR skill already demands spike output; nothing can produce it. This task authors the pair and its artefact contract.

Normative clauses:

1. A new artefact type architectural-spike@1 MUST be defined with frontmatter: spike_id (SPIKE-<task-ID>-<n>), task_id, question (the single decision under investigation), timebox_hours (int, recorded up front), options (array, each {name, hypothesis, evidence[], cost_estimate, risks[]}), recommendation (names exactly one option), confidence (low|medium|high), discarded (array of {name, reason}), created. Body sections: ## Question, ## Options probed, ## Evidence log, ## Recommendation, ## Discard log.
2. architectural-spike-author MUST: take an audited task + repo-context-map as inputs (input envelope), declare its trigger conditions (an ADR is pending AND >= 2 viable options exist, or the task introduces a dependency the repo has never used), enforce the timebox by recording planned vs actual hours, and HALT for the operator when actual exceeds planned by more than 50%.
3. Every evidence[] entry MUST cite something checkable: a file path in the repo, a command + its observed output, or an external URL. Unsupported assertions ("X is faster") without a citation MUST NOT count as evidence.
4. architectural-spike-audit MUST score /10 against a RUBRIC.md with at least these rule families: SPK-STRUCT (frontmatter + sections complete, recommendation names exactly one probed option), SPK-EVID (>= 2 options probed, every option has >= 1 checkable evidence entry, recommendation cites evidence), SPK-BOX (timebox recorded, overrun HALT honored), SPK-DISC (discard log non-empty whenever options were rejected). Only 10/10 passes; the audit report format mirrors REPORT_FORMAT.md conventions of the existing full pairs.
5. architecture-decision-record-author/SKILL.md MUST be updated to name architectural-spike@1 as its spike input (replacing the dangling prose reference), and to state the fallback when no spike exists (lean profile: ADR options table carries the evidence inline).
6. Both skills MUST carry acceptance TRIGGER_TESTS.md with >= 6 cases each (>= 3 positive, >= 3 negative), following the conventions of the existing trigger-test files (TASK-SKILL-112 lineage).
7. File layout MUST match the full-pair convention exactly as listed in new_files, so the TASK-SKILL-118 parity checker passes over this pair with zero exemptions.

§2 - Why this design

The artefact is deliberately evidence-first: the rubric makes uncited claims worthless, which is what separates a spike from an opinion. Timebox recording (plan vs actual, HALT on 1.5x) encodes the whole point of a spike - bounded spend - as a checkable invariant rather than advice. Reusing the full-pair file layout means no new tooling: parity checks, chain-coverage checks, and the plugin build all treat the pair like the existing ten.

§3 - Contract

Input envelope (author): {task_path, repo_context_map_path, question, timebox_hours, output_dir}. Output envelope: {spike_path, spike_id, recommendation, confidence, halted}. Audit input: {spike_path}; audit output: {verdict: pass|fail|needs_human, score, findings[]}. Envelope JSONs are normative in envelopes/.

§4 - Acceptance criteria

1. Artefact contract is complete (§1 #1) - SKILL.md's artefact section defines every frontmatter field with types and the five body sections; a sample spike in the acceptance fixtures parses against it.
2. Trigger conditions are declared and tested (§1 #2, #6) - author TRIGGER_TESTS.md contains >= 6 cases including "ADR pending with 2 options" (fire) and "single obvious option" (no fire).
3. Timebox HALT is normative (§1 #2) - PIPELINE.md contains the plan/actual recording step and the > 1.5x HALT with operator wording; INVARIANTS.md lists it as an invariant.
4. Evidence rule is enforceable (§1 #3, #4) - RUBRIC.md's SPK-EVID rules reject a fixture spike whose option carries only an uncited claim; the fixture and expected finding appear in audit acceptance tests.
5. Rubric families complete, 10/10 gate (§1 #4) - RUBRIC.md defines SPK-STRUCT/EVID/BOX/DISC with rule IDs and states the 10/10 pass bar.
6. ADR input is wired (§1 #5) - architecture-decision-record-author/SKILL.md names architectural-spike@1 and the lean-profile fallback; the dangling prose is gone.
7. Layout parity (§1 #7) - every file in new_files exists; the TASK-SKILL-118 parity check (or, before it lands, a manual ls against the task pair layout) shows no missing file class.

§5 - Verification

Acceptance-driven (skills are contracts, not code):

• modules/skill/architectural-spike-author/acceptance/TRIGGER_TESTS.md - AC 2 cases (>= 6).
• modules/skill/architectural-spike-audit/acceptance/TRIGGER_TESTS.md - AC 4 fixture + expected SPK-EVID finding; plus a 10/10 clean fixture.
• Structural check for AC 1, 3, 5, 6, 7: bash tools/install/check-chain-coverage.sh pair rule (from TASK-SKILL-116) plus grep assertions listed inline in the two TRIGGER_TESTS.md preambles (each acceptance file opens with its own "how to verify this pair" block, executable line by line).

§6 - Implementation skeleton

Author SKILL.md follows the section order of implementation-plan-author (description, triggers, inputs, artefact spec, pipeline pointer, HALT semantics). RUBRIC.md follows task-audit/RUBRIC.md's family/rule-ID formatting. Envelope JSONs mirror the existing pairs' shape (json-schema-lite objects with required arrays).

§7 - Dependencies

None upstream (authoring is self-contained). Blocks TASK-CUO-209 (the pair joins the vendored set there). TASK-SKILL-118's parity checker adopts this pair automatically once both land.

§8 - Example payloads

# spike frontmatter (fixture)
spike_id: SPIKE-TASK-MEMORY-130-1
task_id: TASK-MEMORY-130
question: "MMR chain vs plain hash chain for audit checkpoints?"
timebox_hours: 6
options:
  - name: mmr
    hypothesis: "O(log n) inclusion proofs justify complexity"
    evidence: ["bench: scripts/bench_mmr.sh output 2026-07-12", "modules/memory/cyberos/data/AGENTS.md §6.4"]
    cost_estimate: "3-4 days"
    risks: ["new dependency surface"]
recommendation: mmr
confidence: medium
discarded: [{name: plain-hash, reason: "linear proof size fails NFR-PERF-012"}]

§9 - Open questions

None blocking. Whether ship-tasks gains an optional spike step between steps 2 and 3 is TASK-CUO-209 territory (workflow change), not this pair's.

§10 - Failure modes inventory

1. Spike becomes a design doc (unbounded) - SPK-BOX fails the audit when actual hours are absent or the HALT was skipped; the timebox is data, not vibes.
2. Evidence rot (cited file later deleted) - audit checks citations resolve AT AUDIT TIME; later rot is caught by TASK-SKILL-119's anchor checker class, out of scope here.
3. Recommendation names an unprobed option - SPK-STRUCT explicit rule; fixture covers it.
4. Author invoked with no real alternatives - negative trigger tests pin "single obvious option -> do not fire"; the ADR proceeds without a spike per §1 #5 fallback.
5. Confidence inflation - confidence is an enum the audit cross-checks: high with a single evidence entry per option triggers a SPK-EVID finding.

§11 - Implementation notes

Keep artefact version pinned at @1 in both SKILL.md files; version bumps follow the contracts CHANGELOG discipline. The pair is NOT added to build.sh here - vendoring is TASK-CUO-209's single expansion, kept out of this task to avoid two writers on build.sh:28.

End of TASK-SKILL-117.