Record architecture decisions: add docs/adr from fleet evidence

[cpitzi]

The fleet-wide ADR reconstruction (2026-08-13) recovered this repo's architectural decisions from commit history, issues/PRs, CLAUDE.md, incident reports, and session archives. This issue asks for them to be recorded in-repo as decision records.

Deliverable

• docs/adr/README.md — an index of the entries, opening with a one-paragraph provenance note: reconstructed 2026-08-13 from repo history and fleet records, dates are the original decision dates.
• One docs/adr/0001-<slug>.md per decision below.
• A short "Architecture decisions" line in the main README linking the directory.

Format

Match the existing fleet ADR style (see drosera/docs/adr/0001-*.md / solidago/docs/decisions/0001-*.md): # ADR-NNNN: <title>, **Status:** Accepted (<original date>; reconstructed 2026-08-13), then Context, Decision, Alternatives, Consequences. The Alternatives section must contain (a) the alternatives actually recorded in the evidence, and (b) one or two options explicitly marked "retrospective — not considered at the time", each honestly assessed (worse, better, or lateral — and why). Do not present retrospective options as if they were weighed historically.

Ground rules

• Verify every anchor below (issue/PR numbers, files, dates) against this repo before asserting it in a doc. If something doesn't check out, drop it or state it as uncertain — never copy claims blind.
• Public repo: add no LAN IPs, hostnames, or network topology beyond what this repo already discloses.
• Scope: only the ADR files + the README link. No other changes; adjacent problems get their own issue.
• Open a PR with Closes #<this issue>. Do not arm auto-merge — these PRs wait for human review.

Decisions to record

1. The product is a versioned prompt program. prompts/v<N>-auditor.md, new integer per revision, old versions kept for diffing/reproduction; the multi-model pipeline decomposition designed (docs/architecture.md, Implement pipeline decomposition (Opus / Sonnet / Haiku stages) #11) but deliberately unimplemented. Retrospective: a scripted Crossref/API tool (better determinism; wrong genre — prompt-engineering-as-product is the exhibit, and live web search is the verification surface by design).
2. Paired evaluation sets as regression gates. Adversarial-30 (detection) + the real-article corpus (feat(test-sets): add real-article clean corpus for false-positive calibration #42, false positives); "a revision that flags more on both is just noisier"; the clean corpus caught v5 over-penalization (Scoring formula: D×3 base cost makes the headline score misleading for large reference lists #43: a real article scored ~22, corrected to 96 in v6); baselines committed per version (feat(baseline): establish v5 production baseline — regression gate, H9 functional check #41/feat(baseline): v6 production baseline — H10 functional check, regression gate, scoring calibration #46).
3. Four-tier risk classification over a continuous score; non-accusatory language as a design requirement. "Continuous scores invite gaming and false precision"; verdicts never accuse ("predatory" is not a finding).