docs(adr): seed manifest as durable per-root state

.conductor/registry/workflows/ado-pr.yaml

@@ -618,9 +618,14 @@ agents:
   #      marker → bot reviewer (probably pr_initial_reviewer). Only
   #      treat items under the `**Blocking concerns**` section as
   #      negative; ignore everything else in the bot's body.
-  #   2. Comment author matches the PR author → exclude (PR author
-  #      narrating their own work doesn't count as feedback to themselves).
-  #   3. Otherwise → human reviewer. Use natural-language judgement.
+  #   2. Otherwise → human reviewer. Use natural-language judgement,
+  #      regardless of whether the comment author matches the PR
+  #      author identity. Polyphony PRs are agent-authored under the
+  #      operator's PAT, so author-identity equality is NOT a valid
+  #      proxy for bot-vs-human — the marker is. Self-approval /
+  #      self-waiting-for-author is a platform branch-protection
+  #      concern, not the analyzer's. See
+  #      docs/decisions/pr-feedback-identity-classification.md.
   #
   # Output digest contract: `feedback_digest` is a stable hash of
   # (comment ids + thread ids + vote tuples + body sha) of everything
@@ -690,24 +695,28 @@ agents:
          marker-posting agent). Only treat items under the `**Blocking
          concerns**` section as negative; ignore everything else in
          the bot's body.
-      2. Else if `author == PR author identity` → **author self-comment**.
-         Exclude entirely (the PR author narrating their own work is
-         not feedback to themselves).
-      3. Else → **human reviewer**. Use natural-language judgement of
+      2. Else → **human reviewer**. Use natural-language judgement of
          the comment's sentiment and any embedded code-review concerns.
+         **Do NOT compare `author` against the PR author identity** —
+         polyphony PRs are agent-authored under the operator's PAT, so
+         the operator's own comments on the PR are valid human-in-the-
+         loop interventions and must be treated as feedback.
 
       For reviewer **votes** on ADO: a `-5` (waiting for author) vote
-      counts as negative feedback. A `+5` (approved with suggestions)
-      vote ALONE does not — but if it's paired with unresolved actionable
-      threads, the threads still win and you should treat the unresolved
-      threads as negative feedback. A `+10` or `-10` was already terminated
-      by the deterministic classifier; if you see one here it means
-      mixed signals — defer to the threads / comments to decide.
+      counts as negative feedback regardless of who cast it (including
+      the PR author themselves — self-vote enforcement is a platform
+      branch-protection concern, not yours). A `+5` (approved with
+      suggestions) vote ALONE does not — but if it's paired with
+      unresolved actionable threads, the threads still win and you
+      should treat the unresolved threads as negative feedback. A
+      `+10` or `-10` was already terminated by the deterministic
+      classifier; if you see one here it means mixed signals — defer
+      to the threads / comments to decide.
 
       Unresolved threads (status active / pending — anything that is NOT
       resolved / fixed / closed / by-design / won-fix) that contain at
-      least one in-scope (non-author) comment count as negative feedback
-      even if the comment body is short.
+      least one non-bot comment count as negative feedback even if the
+      comment body is short.
 
       ## Output
 

.conductor/registry/workflows/github-pr.yaml

@@ -491,9 +491,14 @@ agents:
   #      marker → bot reviewer (probably pr_initial_reviewer). Only
   #      treat items under the `**Blocking concerns**` section as
   #      negative; ignore everything else in the bot's body.
-  #   2. Comment author matches the PR author → exclude (PR author
-  #      narrating their own work doesn't count as feedback to themselves).
-  #   3. Otherwise → human reviewer. Use natural-language judgement.
+  #   2. Otherwise → human reviewer. Use natural-language judgement,
+  #      regardless of whether the comment author matches the PR
+  #      author identity. Polyphony PRs are agent-authored under the
+  #      operator's PAT, so author-identity equality is NOT a valid
+  #      proxy for bot-vs-human — the marker is. Self-approval /
+  #      self-changes-requested is a platform branch-protection
+  #      concern, not the analyzer's. See
+  #      docs/decisions/pr-feedback-identity-classification.md.
   #
   # Output digest contract: `feedback_digest` is a stable hash of
   # (comment ids + thread ids + vote tuples + body sha) of everything
@@ -563,23 +568,25 @@ agents:
          marker-posting agent). Only treat items under the `**Blocking
          concerns**` section as negative; ignore everything else in
          the bot's body.
-      2. Else if `author == PR author identity` → **author self-comment**.
-         Exclude entirely (the PR author narrating their own work is
-         not feedback to themselves).
-      3. Else → **human reviewer**. Use natural-language judgement of
+      2. Else → **human reviewer**. Use natural-language judgement of
          the comment's sentiment and any embedded code-review concerns.
-
-      For reviewer **votes**: a `changes_requested` vote from a non-
-      author reviewer counts as negative feedback regardless of any
-      explanatory comment. An `approved` vote with no unresolved
-      threads was already terminated by the deterministic classifier —
-      if you see it here it means the classifier returned `none` due
-      to mixed signals (e.g. approved AND unresolved threads), so the
-      threads still win.
+         **Do NOT compare `author` against the PR author identity** —
+         polyphony PRs are agent-authored under the operator's PAT, so
+         the operator's own comments on the PR are valid human-in-the-
+         loop interventions and must be treated as feedback.
+
+      For reviewer **votes**: a `changes_requested` vote counts as
+      negative feedback regardless of who cast it (including the PR
+      author themselves — self-approval / self-changes-requested is
+      a platform branch-protection concern, not yours). An `approved`
+      vote with no unresolved threads was already terminated by the
+      deterministic classifier — if you see it here it means the
+      classifier returned `none` due to mixed signals (e.g. approved
+      AND unresolved threads), so the threads still win.
 
       Unresolved threads (`is_resolved == false && is_outdated != true`)
-      that contain at least one in-scope (non-author) comment count as
-      negative feedback even if the comment body is short.
+      that contain at least one non-bot comment count as negative
+      feedback even if the comment body is short.
 
       ## Output
 

.conductor/registry/workflows/plan-level.yaml

@@ -417,7 +417,7 @@
       - "plan"
       - "derive-ancestor-chain"
       - "--root-id"
       - "{{ root_resolver.output.resolved_root_id }}"
       - "--item-id"
       - "{{ workflow.input.work_item_id }}"
     routes:
@@ -433,7 +433,7 @@
 
       Could not derive the plan-tree ancestor chain for work item
       **{{ workflow.input.work_item_id }}** under root
       **{{ root_resolver.output.resolved_root_id }}**.
 
       **Error:** {{ ancestor_chain.output.error if ancestor_chain.output.error is defined else "Unknown error" }}
 
@@ -1616,9 +1616,13 @@
   # Identity policy (load-bearing — encoded in the prompt):
   #   1. Comment with a `<!-- polyphony:agent-comment agent=... -->`
   #      marker → bot reviewer. INCLUDE the comment's content.
-  #   2. Comment author matches the PR author → exclude (PR author
-  #      narrating their own work doesn't count as feedback to itself).
-  #   3. Otherwise → human reviewer. INCLUDE.
+  #   2. Otherwise → human reviewer. INCLUDE, regardless of whether
+  #      the comment author matches the PR author identity. Polyphony
+  #      PRs are agent-authored under the operator's PAT, so
+  #      author-identity equality is NOT a valid proxy for bot-vs-
+  #      human — the marker is. Self-approval / self-changes-requested
+  #      is a platform branch-protection concern, not the analyzer's.
+  #      See docs/decisions/pr-feedback-identity-classification.md.
   #
   # Severity policy (also load-bearing):
   #   - For polyphony's own bot reviewers (e.g. plan_reviewer), only
@@ -1700,20 +1704,23 @@
          posting agent). Only treat items under the `**Blocking
          concerns**` section as negative; ignore everything else in the
          bot's body.
-      2. Else if `author == PR author identity` → **author self-comment**.
-         Exclude entirely (the PR author narrating their own work is
-         not feedback to themselves).
-      3. Else → **human reviewer**. Use natural-language judgement of
+      2. Else → **human reviewer**. Use natural-language judgement of
          the comment's sentiment and any embedded code-review concerns.
-
-      For reviewer **votes**: a `changes_requested` vote from a non-
-      author reviewer counts as negative feedback regardless of any
-      explanatory comment. A `rejected` vote was already terminated by
-      the deterministic classifier — you should not see it here.
+         **Do NOT compare `author` against the PR author identity** —
+         polyphony PRs are agent-authored under the operator's PAT, so
+         the operator's own comments on the PR are valid human-in-the-
+         loop interventions and must be treated as feedback.
+
+      For reviewer **votes**: a `changes_requested` vote counts as
+      negative feedback regardless of who cast it (including the PR
+      author themselves — self-vote enforcement is a platform branch-
+      protection concern, not yours). A `rejected` vote was already
+      terminated by the deterministic classifier — you should not see
+      it here.
 
       Unresolved threads (`is_resolved == false && is_outdated != true`)
-      that contain at least one in-scope (non-author) comment count as
-      negative feedback even if the comment body is short.
+      that contain at least one non-bot comment count as negative
+      feedback even if the comment body is short.
 
       ## Output
 

docs/decisions/pr-feedback-identity-classification.md

@@ -0,0 +1,128 @@
+# PR Feedback Identity Classification — Marker, Not Author Identity
+
+> **Status:** Accepted (2026-05-25).
+> **Affects:** `.conductor/registry/workflows/{plan-level,github-pr,ado-pr}.yaml`
+> — specifically the `pr_feedback_analyzer` prompt in each.
+> **Supersedes:** the prior "identity policy" rule that excluded any
+> comment whose author matched the PR author identity.
+
+## Context
+
+Three workflows (`plan-level.yaml`, `github-pr.yaml`, `ado-pr.yaml`) run a
+sentiment-driven review loop centered on a `pr_feedback_analyzer` agent.
+The analyzer reads the poll envelope (comments, threads, votes) and
+returns `has_negative_feedback` to drive the revise/merge/abort routing.
+
+Polyphony's PRs are almost always **agent-authored under the operator's
+PAT** — the coder agent (and the marker-posting reviewer agents) push
+through the same token the human operator uses. As a result, the PR
+`author_identity` on the platform API is the **human operator**, not
+the agent that actually drove the commits or posted the comments.
+
+To disambiguate bot-from-human under a shared identity, polyphony emits a
+deterministic HTML marker on every machine-posted comment:
+
+```html
+<!-- polyphony:agent-comment agent=plan_reviewer head_sha=abc1234 run_id=xyz -->
+```
+
+The parser lives at `src/Polyphony/Commands/PrCommentMarker.cs` and is
+called out explicitly in its summary as existing for "the common ADO case
+where `plan_reviewer_poster_ado` shares the operator's PAT".
+
+### The bug this ADR addresses
+
+Through 2026-05, the analyzer prompt carried a three-tier classifier:
+
+1. Has `polyphony:agent-comment` marker → **bot**.
+2. Else if `author == PR author identity` → **author self-comment**, exclude.
+3. Else → **human reviewer**.
+
+Rule 2 is wrong. It excludes the human operator's own comments on
+agent-authored PRs — exactly the highest-priority signal in the system,
+the **human-in-the-loop intervention**. An incident on 2026-05-25
+surfaced this: an analyzer run reported `has_negative_feedback: false`
+while three review threads from the operator identity were silently
+dropped under rule 2.
+
+The stated rationale ("the PR author narrating their own work doesn't
+count as feedback to themselves") conflates two distinct cases:
+
+- The **agent** narrating its own progress — yes, noise, but the marker
+  already identifies these as bot comments.
+- The **human operator** intervening on an agent-authored PR — this is
+  authoritative feedback that must drive the revise loop.
+
+Identity-equality is not a valid proxy for "narrator vs. intervener"
+when the agent and the operator share an identity. The marker is.
+
+## Decision
+
+The analyzer classifies comments **by marker presence, not by author
+identity**. The full rule, applied uniformly to comments AND votes:
+
+1. **Marker present** → bot. Classify by the marker's `agent`
+   attribute. Apply existing bot logic (e.g. the
+   `**Blocking concerns**` section convention).
+2. **Marker absent** → human. Apply natural-language judgement, regardless
+   of whether the author identity matches the PR author identity.
+
+There is **no identity comparison** anywhere in the rule. A `+10` /
+`approved` vote from the operator on their own PR counts as positive
+feedback. A `-5` / `changes_requested` vote from the operator counts as
+negative feedback. A free-form comment from the operator is judged on
+its content like any other human comment.
+
+### Why no carve-out for votes
+
+A previous draft of this ADR kept identity-exclusion for the **votes**
+case on the grounds that "self-approval is platform-blocked anyway".
+That carve-out was rejected because:
+
+- Self-approval is only platform-blocked at the **final merge to main**
+  in default branch-protection configurations. Earlier-stage approvals
+  (intermediate gates, draft reviews, ADO `+5` "approved with
+  suggestions" on a non-protected branch) are not universally blocked.
+- Whether self-approval is permitted at any given stage is a
+  **policy/branch-protection concern enforced by the platform**, not a
+  concern of the polyphony feedback analyzer. If the platform blocks the
+  merge for policy reasons, the platform surfaces that downstream
+  (`mergeable_state`, merge-time refusal). The analyzer's job is only to
+  classify the **feedback content**, not to predict or enforce platform
+  merge policy.
+
+Mixing platform-policy reasoning into the analyzer was the original
+mistake and the carve-out would have replicated it in miniature.
+
+## Consequences
+
+- **Human operator interventions on agent-authored PRs now drive the
+  revise loop.** This is the intended human-in-the-loop affordance and
+  the load-bearing reason for the change.
+- **Self-approval is honored as positive signal at the analyzer layer.**
+  Whether the resulting merge actually proceeds is the platform's call
+  via branch protection / approval-required policy. Operators who want
+  to forbid self-approval should configure that in branch protection,
+  not rely on the analyzer to silently filter it out.
+- **The `author_identity` field stays in the rendered prompt context**
+  for diagnostic value (it appears in the "Context — Poll envelope"
+  block), but the analyzer no longer compares against it. Future
+  versions may drop the field entirely; keeping it for now eases
+  triage of analyzer reasoning traces.
+- **The structural consistency lint
+  (`lint-sentiment-loop-consistency.ps1`) is unaffected** — it pins
+  agent shape, output schema, and route presence, not prompt text.
+- **The marker parser
+  (`src/Polyphony/Commands/PrCommentMarker.cs`) becomes more
+  load-bearing.** Any future bot poster that fails to inject the marker
+  will have its comments classified as human feedback. The existing
+  Pester tests around the marker remain the safety net.
+
+## Non-goals
+
+- This ADR does not change which votes are **terminal** (handled by the
+  deterministic classifier upstream of the analyzer). The terminal
+  classifier's identity rules — if any — are out of scope and continue
+  to be evaluated on their own merits.
+- This ADR does not change marker injection by polyphony's bot posters;
+  it only changes how the analyzer interprets the absence of a marker.