feat(discovery): family-aware/target-as-LLM hunt class + mid-loop early-warning + attack-chain rollup (T3MP3ST R5/R8/R15; v0.26.0)

[omkhar]

Closes #160. PR-C of 3 (final) implementing the T3MP3ST-audit survivors. Tier-3, prose-only reference additions — no validator or canonical-body change.

What changed

• R5 — family-aware target-class vocabulary + target-as-LLM hunt class. Before lane routing, characterize the target's trust-boundary paradigm (memory/process, state/invariant, model/guardrail, network-protocol) and instantiate the existing lane model + bug-class questions with family-appropriate vocabulary (subsystem_inventory.other + the open bug-class list are the sanctioned extension points — no enum/validator change). Adds a Model/guardrail (target-as-LLM) analyst lane for when the target itself is an LLM/agent app, explicitly distinct from the analyzer's own XPIA/tool-firewall containment. Candidate-labeling recall only; every candidate still passes the normal gates. (discovery-specialist-router.md + discovery-orchestrator.md)
• R8 — non-blocking mid-loop early-warning. The Critique/Adapt stage surfaces evidence-without-hypothesis and closure-needs-retest sooner than the terminal drift gate, scheduling the missing step into next_commands; non-blocking, never auto-dispositions, and never substitutes for drift-gate.md (the single source of truth). (loop-engineering.md)
• R15 — per-asset attack-chain rollup. An optional, ungated presentation table grouping several validated findings on one asset into an ordered chain rated by worst-step = max already-proven demonstrated_impact — no new chain-severity synthesis, no new proof; discovered stays distinct from proven. (artifact-index.md)

Process

• Re-scoped survivors of the loop-until-dry adversarial review (Tier-3).
• A three-lens /simplify pass (reuse/dedup, simplification, altitude) confirmed all additions are genuinely new or correctly-cited pointers, verified R8's non-blocking/defers-to-drift-gate and R15's optional/no-synthesis framing, and drove three concision cuts.
• No canonical-body edit (trusted-body digest unchanged), no validator/contract change. Reference ceilings raised (router 85→98, orchestrator 104→107, loop-engineering 112→122, artifact-index 236→246), consensus in reviews/discovery-targeting-reporting-consensus.md; five-agent docs recheck + agent-feature-matrix.md refresh in docs/authoritative-sources.md.
• make sync + make verify clean (686 unit + 9 integration).

🤖 Generated with Claude Code

[omkhar]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 24ff212b68

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[omkhar]

Addressed (real): demonstrated_impact is an open capability description, not an ordered enum, so a max() over it is undefined and invites exactly the synthesis the clause forbids. The optional rollup now lists each step's own already-proven demonstrated_impact and severity per row, with no computed chain rating / no max-reduce. CHANGELOG + consensus updated to match.

[omkhar]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Breezy!

Reviewed commit: 55961ca50f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".