feat(router): unify frontend and standalone selection core

.agents/skills/author-baseline-dgd/SKILL.md

@@ -64,9 +64,13 @@ prefer aggregated unless the user's SLOs demand otherwise).
    workloads; raise TP above `min_tp` only when headroom demands it, recording the replica cost).
 2. **Choose topology conservatively**: an aggregated single-node layout unless the user's hardware or SLOs force
    otherwise. The baseline's job is to run and measure, not to win; the optimization loop owns improvement.
-3. **Scaffold from the nearest recipe**: copy its structure (components, probes, service wiring, image versions
-   for the chosen backend) and replace model, parallelism, resources, and any hardware-bound fields, naming every
-   replacement. Never carry a hardware-bound topology, transport, or checkpoint choice across without evidence it
+3. **Scaffold from the nearest recipe**: copy its structure (components, probes, service wiring) and replace
+   model, parallelism, resources, and any hardware-bound fields, naming every replacement. Image versions are NOT
+   copied blindly: when a recipe dossier snapshot exists (`<EXP_ROOT>/analysis/recipe-dossier/`), read the
+   scaffold candidate's recorded verdict and failed conditions, and never carry forward an image the dossier
+   marked mutable, unresolved, stale, or quarantined (Tier 4); choose an image that resolves to an immutable
+   release tag or digest for the chosen backend and record the resolution in the evidence table. A draft whose
+   image fails that gate is not ready for confirmation. Never carry a hardware-bound topology, transport, or checkpoint choice across without evidence it
    fits the target. A manifest expresses REQUIREMENTS (GPU type, count, memory), never observed cluster state:
    do not pin node names or encode which nodes happen to be free, and preserve the recipe's scheduling
    MECHANISMS (tolerations, product-label node selectors) while retargeting their VALUES to the contract's

.agents/skills/consult-perf-knowledge/SKILL.md

@@ -178,7 +178,14 @@ when current evidence shows that family dominates the objective or a broad scan
 Follow `agent-docs/guides/knob-tuning/tuning-hierarchy.md`:
 
 1. Classify the exact model and compute memory fit, minimum parallelism, and headroom.
-2. Complete the exploration-versus-exploitation calibration and full broad-lever scan above.
+2. Complete the exploration-versus-exploitation calibration and full broad-lever scan above. As part of the broad
+   scan, invoke `find-serving-recipe` once per engagement, reusing the latest snapshot listed in
+   `EXP_ROOT/analysis/recipe-dossier/index.md` on later iterations when one exists, and record the snapshot's
+   path and SHA256 in `knowledge-consult.md`, so that every
+   recipe-shaped candidate under consideration carries a provenance verdict: a `deployable` or `hypothesis` grade
+   candidate enters the lever shortlist with its dossier entry attached, and a `ceiling-only` candidate may inform
+   expected-performance headroom but is never shortlisted for deployment. A recipe-sourced candidate goes through
+   this same selection and the adversarial review like any other hypothesis; it never replaces the baseline.
 3. Consider topology first as a hypothesis category per the tuning hierarchy; an inherited layout is a candidate,
    not a settled decision.
 4. When topology is viable, consider Tier 2 families in default priority order, then finish screening the remaining

.agents/skills/report-skillpack-issue/SKILL.md

@@ -0,0 +1,188 @@
+---
+name: report-skillpack-issue
+description: Reports a defect in the optimization skillpack itself (the skills under .agents/skills/, the documents under agent-docs/, the role contracts under agents/, and the repository AGENTS.md files) as a GitHub issue on ai-dynamo/dynamo, using the repository's agent-reported issue conventions. Use when a pack rule contradicts another pack rule, a cross-reference points at a file or section that does not exist, an instruction cannot be executed as written, a factual claim about a tool or flag is wrong, or a rule repeatedly fights the observed environment. Do not use for bugs in Dynamo itself, for engagement-specific problems, or to request new features.
+license: Apache-2.0
+user-invocable: true
+metadata:
+  author: NVIDIA
+  tags:
+    - dynamo
+    - skillpack
+    - telemetry
+    - github
+---
+
+# Skill: Report a Skillpack Defect
+
+<!--
+SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+SPDX-License-Identifier: Apache-2.0
+-->
+
+Issues filed against the skillpack are the maintainers' primary field telemetry. A defect you hit and route around
+silently gets hit by every other user's agent too. Filing takes two minutes; the report is more valuable than the
+workaround.
+
+This skill follows the repository's agent-reported issue conventions
+(`.github/ISSUE_TEMPLATE/agent-reported.yml`): the `[AGENT]: ` title prefix, the `agent-reported` label, a
+declared agent identity, and at most one issue per session.
+
+## Step 1: Confirm the defect is in the pack, not the environment
+
+Re-read the suspect rule or skill file in full before reporting. Classify the defect:
+
+- `contradiction` — two pack statements cannot both be followed.
+- `dead-reference` — a cross-reference targets a file, section, or field that does not exist.
+- `unexecutable` — an instruction cannot be carried out as written (missing input, impossible ordering, undefined
+  term with no declared source).
+- `factual-error` — a claim about a tool, flag, default, or format is wrong; verify against the tool's own
+  source or help output first, and quote that verification in the report.
+- `environment-mismatch` — the rule assumes something (paths, permissions, cluster shape) that the target
+  environment class does not satisfy; report only when the mismatch looks general, not site-specific.
+- `missing-coverage` — the pack gives no guidance for a component or situation it plainly should cover.
+
+If the problem disappears on a careful re-read, or is specific to one site's configuration, do not file.
+
+## Step 2: Record your agent identity and the exact pack version
+
+The report must state the driver model and the skills commit it was running (for example
+`claude-opus-5, skills @ abc1234`):
+
+```bash
+git rev-parse --short HEAD
+```
+
+Run that from the repository root. In every command in this skill, replace placeholders BEFORE running
+and never leave angle brackets in a shell line: `<...>` parses as redirection and can truncate files.
+
+If the pack was vendored without git history, record the release or image tag it came from. A report that cannot
+say which version it observed cannot be triaged.
+
+## Step 3: Sanitize
+
+The issue lands in a public repository. Nothing you emit, in the title, the body, a comment, a search term,
+or a version string, may contain: Quote pack text from the public repository freely (not from a locally modified or vendored copy, which may
+carry private details); describe your environment only in generic terms (GPU class, backend, harness). When in doubt, leave it out — the rule file,
+version, and defect class are usually enough to reproduce.
+
+## Step 4: Check for an existing report
+
+Search titles AND bodies in one bounded, body-inclusive query. A labeled query adds nothing (`--label`
+ANDs into the search, so it is a strict subset of the unlabeled one), the default limit of 30 hides real
+matches for common keys such as `README.md`, and `--json` returns bodies inline instead of costing one
+`gh issue view` per hit. Use the bare filename as the search term (no backticks):
+
+```bash
+gh issue list --repo ai-dynamo/dynamo --state all --limit 100 --json number,title,body \
+  --search 'benchmark-isolation.md in:title,body'
+```
+
+Review every hit whose title starts with `[AGENT]: ` or whose body mentions the same file and
+defect class, and read each candidate's comments too, because confirmations of an existing defect
+are filed as comments (below) and do not appear in title or body search:
+
+```bash
+ISSUE_NUMBER=123
+gh issue view "$ISSUE_NUMBER" --repo ai-dynamo/dynamo --comments
+```
+
+If a matching issue exists, do not file a duplicate: draft the same body (step 5), route it through
+step 6 with the target recorded as "comment on #123", and on approval run the command with BOTH
+variables assigned in the same shell invocation (nothing persists between harness shell calls):
+
+```bash
+ISSUE_NUMBER=123
+BODY_FILE=$EXP_ROOT/analysis/skillpack-issue-drafts/001-perf-analyzer-body.md
+gh issue comment "$ISSUE_NUMBER" --repo ai-dynamo/dynamo --body-file "$BODY_FILE"
+```
+
+File at most one new issue per session; if the session surfaced several defects, put the most
+impactful one in the issue and list the rest briefly in its body.
+
+## Step 5: Draft the issue
+
+Title: `[AGENT]: <file path relative to repo root>: <one-line defect>`.
+
+Write the body with a file-writing tool, never by echoing it through a shell, to a run-scoped path that
+names the invoking role, because several roles can hit defects in the same engagement window: inside an
+engagement `<EXP_ROOT>/analysis/skillpack-issue-drafts/NNN-<role>-body.md` (NNN increasing), otherwise
+`./skillpack-issue-drafts/NNN-<role>-body.md` under the current working directory. Never use a shared
+fixed path such as `/tmp`, where another role's write can replace an approved body before it is
+submitted. Delete the draft file once its issue or comment is filed, so a later invocation cannot inherit
+a stale body. Follow the
+agent-reported template's structure:
+
+```markdown
+### Agent identity
+
+<driver model>, skills @ <commit or tag>
+
+### What the instructions said vs what you verified
+
+Defect class: <contradiction | dead-reference | unexecutable | factual-error | environment-mismatch | missing-coverage>
+Location: <file path and the section heading or quoted sentence>
+
+<exact quote(s) of what the pack says; for a contradiction, quote both sides>
+
+<the generic, sanitized observation that contradicts it, how it was verified, and what the agent could not do,
+did wrong, or had to route around>
+
+### Suggested correction
+
+<the wording you would have needed; omit the section if unsure>
+```
+
+## Step 6: Who files, and how
+
+Filing a public issue or comment is an external side effect that needs the operator's approval of the
+exact drafted text. Two cases:
+
+- **Dispatched sub-role** (interviewer, deployer, analyzer, generator, challenger): you do not talk to
+  the operator, so you never run `gh issue create` or `gh issue comment`. Record the draft in
+  `<EXP_ROOT>/analysis/skillpack-defects.md` (dated section, defect class, location, draft title,
+  draft body path, proposed target: new issue or comment on #N, status `unfiled`) and return that
+  path to your parent. The top-level agent batches every unfiled draft, presents them to the operator,
+  and on approval files at most ONE new issue per session (secondary defects go in its body) plus any
+  approved comments, then updates each draft's status to `filed` with the URL. This is what makes
+  "one issue per session" hold across roles.
+- **Standalone invocation by an operator**: the operator is present; show the complete title and body,
+  file only on approval.
+
+Whoever files: assign the title from a quoted heredoc in the same shell invocation as the command
+(harness shells do not persist variables between calls), and pass the approved body file unchanged.
+Quoting `"$ISSUE_TITLE"` at the call site protects nothing on its own: backticks and `$()` in a drafted
+title are interpreted when the variable is *assigned*, and the pack's own style backticks every flag.
+Do not edit the body between approval and filing; file the same path the operator saw.
+
+```bash
+BODY_FILE=$EXP_ROOT/analysis/skillpack-issue-drafts/001-perf-analyzer-body.md
+ISSUE_TITLE=$(cat <<'EOF'
+[AGENT]: agent-docs/rules/benchmarking/benchmark-isolation.md: one-line defect
+EOF
+)
+gh issue create --repo ai-dynamo/dynamo --title "$ISSUE_TITLE" --body-file "$BODY_FILE" --label agent-reported
+```
+
+Labels behave two ways depending on the client path: the REST API silently drops labels the
+reporter may not set ("Only users with push access can set labels for new issues. Labels are silently
+dropped otherwise."), while `gh` may instead fail before creating the issue when it cannot resolve a
+label. Handle both: if the command fails on the label, retry without `--label`; if it succeeds,
+verify:
+
+```bash
+ISSUE_NUMBER=123
+gh issue view "$ISSUE_NUMBER" --repo ai-dynamo/dynamo --json labels --jq '[.labels[].name]'
+```
+
+If `agent-reported` is absent, tell the operator so a maintainer can add it; maintainers also triage
+by the `[AGENT]: ` title prefix, and the step 4 search covers unlabeled reports.
+
+## Fallback: no GitHub access or no approval
+
+Write the complete draft to an append-only artifact the invoking role owns and tell the operator
+where it is. Inside an optimization engagement that is `<EXP_ROOT>/analysis/skillpack-defects.md`
+(one dated section per draft, shaped like an `asks.jsonl` entry: defect class, location, draft
+title, draft body, status `unfiled`); never write under `final/`, whose three artifacts belong to
+`hypothesis-generator` and are written only at stop-request time. Outside an engagement, write
+`skillpack-issue-draft-<n>.md` at the root of the current working directory. A
+drafted-but-unfiled report is still telemetry; a silent workaround is not.

.dockerignore

@@ -25,13 +25,15 @@
 **/*.libtorch_model_store/*
 **/.git
 **/.github
-# Planner test images exercise the version-stamping script without copying the
-# rest of the GitHub metadata into every Docker build context.
+# Test images exercise these scripts without copying the rest of the GitHub
+# metadata into every Docker build context.
 !.github
 .github/*
 !.github/scripts
 .github/scripts/*
 !.github/scripts/apply_dev_version.py
+!.github/scripts/retry_kubectl.sh
+!.github/scripts/collect_vcluster_diagnostics.sh
 **/*backup*/
 .dockerignore
 **/target/*