docs: rescue v1.0-design-final.md into repo

[suzuke]

Why

Phase A reverse-engineering review (5-backend architectural diagnosis) surfaced that the entire v1.0 redesign sprint (M1a → M3) anchors on docs/v1.0-design-final.md as source-of-truth, but the file was never committed to the repo. References are everywhere:

• Module docstrings: strategy.py:9, ledger.py:1-20, sealing.py:7-11, agents/cli_subscription_backend.py:6, locking.py:1-32, etc. all cite per v1.0-design-final.md §X.Y
• Demo gates: docs/M1B-DEMO-GATE.md:4 and docs/M2-DEMO-GATE.md:4 both say "Spec reference: docs/v1.0-design-final.md §M{1b,2} deliverable demo gate"
• PR descriptions: PR feat(m1a): TrialLedger + dual-write + HTML postmortem reporter #2 (M1a), PR feat(m1b): SearchStrategy Protocol + BFTS-lite + Evaluator trust boundary #3 (M1b), PR feat(m3): SubscriptionCLIBackend (CLI experimental) + ClaudeCodeCLIAdapter #10 (M3 cli-subscription), … all claim spec alignment per §INV-1..4, §3.2/§3.3, §4.1, §11
• Sprint vocabulary: §M1a, §M1b, §M2, §M3, §INV-1, §INV-2, §INV-3, §INV-4, §3.2, §3.3, §4.1, §11

A find . -name 'v1.0-design*' only returns the two demo gate files — the spec itself lives only in the author's local working dir.

Effect: every claim of "per spec §X.Y" is currently unverifiable by anyone except the author, and contributors / external reviewers cannot ground architectural debates in a tracked baseline.

What

Imports v1.0-design-final.md verbatim from the author's local notes (~/Documents/Hack/autocrucible-design-v1/, mirrored via /tmp/ac-v1.0-design-final.md).

• 385 lines, dated 2026-04-25 (matches the M1a sprint kickoff commits on Apr 25-27)
• Status: APPROVED FOR IMPLEMENTATION (v3.2, with 4 small invariants folded in)
• Authored by suzuke; reviewed by codex (crucible-reviewer)
• Contains: executive summary, v3 → v3.2 decision log, §INV-1..4, §M1a/M1b/M2/M3 deliverables, §3.x (CLI subscription), §4.x (AttemptNode / EvalResult), §11 (concurrency invariants)

No code or other doc changes. Single-file additive PR.

Test plan

• Confirm the imported content is the intended v3.2 version (no stale draft)
• Decide canonical location:
  • Option A: docs/v1.0-design-final.md (current — flat next to README / CHANGELOG)
  • Option B: docs/spec/v1.0-design-final.md (new subdir if more specs are planned)
• Decide whether to add a docs/README.md index entry pointing readers at this spec
• Optional: cross-link from CLAUDE.md so AI contributors hit it on first read
• Verify references in module docstrings still parse correctly after the file lands (no path changes needed since they all say docs/v1.0-design-final.md)

Risks

• None functional — pure docs addition
• Spec contains historical decision-trail (v3 → v3.1 → v3.2 with reviewer verdicts); some readers may want a cleaner final-only version, but the trail is part of the value (matches PR description conventions used in M1a-M3)

🤖 Generated with Claude Code