diff --git a/.ciagent/PERSONAS.md b/.ciagent/PERSONAS.md
index 1bda1d1..2139fe2 100644
--- a/.ciagent/PERSONAS.md
+++ b/.ciagent/PERSONAS.md
@@ -1,32 +1,174 @@
 ---
 project: acdl
-milestone: v1.22
+milestone: v1.23
 generated_at: 2026-08-11
 generator: lead-developer
 verification_toolchain:
-  typecheck: "python3 -m py_compile tests/test_slides_pipeline.py"
-  test: "pytest tests/test_slides_pipeline.py   # REQ-254..262"
-  build: "bash scripts/render_slides.sh nova-autonomous-cloud-delivery   # HTML + PPTX"
+  typecheck: "python3 -m py_compile scripts/inline_images.py scripts/render_pptx.py tests/test_slides_pipeline.py"
+  test: "pytest tests/test_slides_pipeline.py   # REQ-263..275"
+  build: "bash scripts/render_slides.sh nova-autonomous-cloud-delivery   # HTML + Marp PPTX; then python3 scripts/render_pptx.py   # structured PPTX"
   note: |
-    v1.22 is the Nova Deck Layout Fix — a docs-only NFR milestone. Two
-    active personas: lead-developer (theme CSS + deck markdown + talking
-    points + README + .ciagent metadata), backend-engineer (render scripts
-    + tests). frontend-engineer stays deactivated (decks are markdown =
-    lead-developer territory, per v1.17/v1.18 precedent). No data-engineer
-    (no schema/DB changes). No new personas (the work is CSS + bash +
-    markdown + pytest, all within the two active personas' range).
+    v1.23 is the Nova Deck Cleanup & Python PPTX — a docs/render/test
+    NFR milestone. Two active personas: lead-developer (deck markdown
+    consolidation + inline Marp style CSS + README + .ciagent metadata),
+    backend-engineer (Python scripts inline_images.py + render_pptx.py
+    via python-pptx + render_slides.sh updates + tests + CI YAML +
+    attach_release_asset.py extension). frontend-engineer stays
+    deactivated (decks are markdown + Marp CSS = lead-developer
+    territory, per v1.17/v1.18/v1.22 precedent). No data-engineer (no
+    schema/DB changes). No new personas (the work splits cleanly into
+    narrative+CSS+docs and Python+bash+tests+CI).
 ---
 
-# ACDL — Persona Roster (v1.22 Nova Deck Layout Fix)
+# ACDL — Persona Roster (v1.23 Nova Deck Cleanup & Python PPTX)
 
-> v1.22 roster. Two active personas + one deactivated. This is a docs-
-> only NFR milestone: the work is theme CSS, render scripts, mermaid
-> diagrams, deck markdown, and tests. frontend-engineer stays
-> deactivated (decks are markdown = lead-developer territory, per
-> v1.17/v1.18 precedent). No data-engineer (no schema/DB changes).
+> v1.23 roster. Two active personas + two deactivated. This is a
+> docs/render/test NFR milestone: the work is deck markdown
+> consolidation, Marp inline `style:` CSS (reverting from the standalone
+> `nova-sp-theme.css` to `theme: default` + inline block), two Python
+> render scripts (`inline_images.py` stdlib image inlining +
+> `render_pptx.py` structured PPTX via python-pptx), test updates, CI
+> YAML, README rewrite, and the `attach_release_asset.py` extension to
+> dual-PPTX. frontend-engineer stays deactivated (Marp CSS is a static
+> stylesheet, not a React/Next.js component system — D-148 precedent).
+> No data-engineer (no schema/DB/ORM changes).
 
 ## Active personas
 
+### lead-developer
+- **Domain:** coordination + deck content + inline CSS
+- **Active:** true
+- **Phase-specific:** false
+- **Frameworks:** [] (no framework — owns process + narrative + Marp inline CSS + markdown + README)
+- **Constraints:** ["pragmatic", "battle-tested defaults", "no fabrication (NORTH_STAR honesty model)", "do not change the 4-beat arc", "do not re-introduce badges/version/internal citations", "port the reference deck's exact inline style CSS (the clean look)"]
+- **Territory:**
+  - `docs/presentations/nova-autonomous-cloud-delivery-marp.md` (REQ-263..275 — deck consolidation: merge plain .md into -marp.md, trim word count, remove "penetrate")
+  - `docs/presentations/nova-autonomous-cloud-delivery.md` (DELETE — consolidated into -marp.md)
+  - `docs/presentations/nova-autonomous-cloud-delivery-talking-points.md` (sync to final slide structure)
+  - `docs/presentations/README.md` (rewrite 4-step → 3-step process for single-document + dual-PPTX pipeline)
+  - `docs/presentations/assets/nova-sp-theme.css` (RETIRE from render; keep as reference file)
+  - `.ciagent/**` (PROJECT, ROADMAP, REQUIREMENTS, RESEARCH, PLAN, GRILL, PERSONAS, REVIEW, CHECKPOINT)
+- **Reason:** Owns the deck markdown consolidation (plain `.md` +
+  `-marp.md` → single `-marp.md` with speaker notes as HTML comments),
+  the inline `style:` CSS block (porting the reference deck's exact
+  CSS — the "clean look"), the word-count trim, the "penetrate"
+  removal, the talking-points sync, the README rewrite, and all
+  CIAgent metadata. Is the only persona that touches `.ciagent/**` and
+  the deck markdown/CSS. The inline `style:` block is a Marp
+  frontmatter stylesheet, not a frontend component system (D-148
+  precedent from v1.22).
+- **Phase-specific flag:** none (active for all of P0–P-final).
+
+### backend-engineer
+- **Domain:** Python render scripts + bash + tests + CI
+- **Active:** true
+- **Phase-specific:** false
+- **Frameworks:** ["python-pptx", "bash", "pytest", "marp-cli", "mermaid-cli", "base64/re/mimetypes (stdlib)"]
+- **Constraints:** ["pin CLI versions (no @latest in workflows-src/slides.yml — currently unpinned, must fix)", "python-pptx>=0.6.23 minimum", "stdlib-only for inline_images.py (no external image lib)", "tests must catch the theme-default + inline-style regression (the gap that let v1.22's standalone-theme drift through)", "no raw curl with shell-env tokens"]
+- **Territory:**
+  - `scripts/inline_images.py` (NEW — stdlib base64 image inlining into rendered HTML)
+  - `scripts/render_pptx.py` (NEW — structured editable PPTX via python-pptx)
+  - `scripts/render_slides.sh` (update: drop --theme nova-sp-theme.css; add inline_images.py + render_pptx.py invocations)
+  - `scripts/attach_release_asset.py` (extend to accept multiple file paths: both .pptx + -python.pptx)
+  - `tests/test_slides_pipeline.py` (REQ-262 successor — invert theme tests, retarget CSS tests to inline style block, delete source-md tests, update slide count)
+  - `workflows-src/slides.yml` (add python-pptx install step; pin @latest → @4.5.0/@11.16.0; add -python.pptx to commit list)
+  - `pyproject.toml` (add python-pptx>=0.6.23 to [project.optional-dependencies] test)
+- **Reason:** Owns the Python render scripts (`inline_images.py` is
+  stdlib regex + base64; `render_pptx.py` is python-pptx — a backend
+  Python library, not a frontend framework). The render_slides.sh
+  updates (drop `--theme`, add the two Python script invocations) and
+  the CI YAML (add python-pptx install, pin versions, add
+  -python.pptx to commit list) are backend/scripting tasks. The test
+  suite updates (invert the theme-default test, retarget CSS tests to
+  the inline `style:` block, delete the source-md tests) are the gap
+  that let v1.22's standalone-theme drift through — backend-engineer
+  owns closing it. python-pptx is new to v1.23; backend-engineer's
+  frameworks list gains it.
+- **Phase-specific flag:** none (active for P2 scripts, P5 tests; light
+  touch on P0/P-final).
+
+## Deactivated personas
+
+### frontend-engineer
+- **Active:** false
+- **Domain:** frontend
+- **Frameworks:** ["react", "next.js"] (inert — no territory)
+- **Constraints:** ["component-first", "server-components", "minimal-client-js"] (inert)
+- **Territory:** [] (no territory in v1.23)
+- **Reason:** v1.23 has no frontend; decks are markdown + Marp inline
+  CSS (lead-developer territory); deactivated per PERSONAS.md
+  v1.17/v1.18/v1.22 precedent. The inline `style:` block is a Marp
+  frontmatter stylesheet (CSS rules in the markdown frontmatter), not
+  a React/Next.js component system — it stays lead-developer
+  territory. No reactivation trigger.
+
+### data-engineer
+- **Active:** false
+- **Domain:** data
+- **Frameworks:** [] (inert)
+- **Constraints:** [] (inert)
+- **Territory:** [] (no territory in v1.23)
+- **Reason:** v1.23 has no schema/DB/ORM changes. The milestone is
+  docs + Python scripts + bash + tests + CI only. No reactivation
+  trigger.
+
+## Roster decisions
+
+### D-163 (0.95): Inline Marp `style:` block is lead-developer territory, not frontend-engineer
+The inline `style:` block in the Marp frontmatter is a static CSS
+stylesheet (a block of CSS rules in the markdown frontmatter, applied
+by Marp to the rendered slides), not a React/Next.js component system.
+The v1.17/v1.18/v1.22 precedent (decks are markdown = lead-developer
+territory; the theme CSS is a Marp stylesheet, not a frontend framework
+— D-148) extends to the inline `style:` block. frontend-engineer's
+frameworks (react, next.js) are irrelevant to Marp frontmatter CSS.
+**Decision:** inline `style:` block stays lead-developer territory.
+Confidence 0.95 — the only counter-argument is that CSS is "frontend,"
+but Marp frontmatter CSS is a static stylesheet authored in the
+markdown source, not a component system.
+
+### D-164 (0.90): No new personas for v1.23
+The work is markdown + inline CSS + Python scripts + bash + tests + CI
+YAML + README. All of this is within the two active personas' range
+(lead-developer: markdown + CSS + README + metadata; backend-engineer:
+Python + bash + tests + CI). Creating a separate "slides-engineer" or
+"pptx-engineer" persona would fragment ownership of the deck markdown
+(lead) and the render scripts (backend). The python-pptx work is
+unambiguously backend (a Python library + script), even though the
+slide layout decisions within `render_pptx.py` are co-owned with lead-
+developer. **Decision:** no new personas. Confidence 0.90 — follows
+v1.22 D-149 precedent.
+
+### Territory-overlap resolution (co-ownership)
+
+| Path | Primary | Co-owner | Why |
+|------|---------|----------|-----|
+| `scripts/render_pptx.py` | backend-engineer (Python code + python-pptx API) | lead-developer (slide layout decisions: which image where, font sizes, bullet structure) | The script is backend (Python); the slide-by-slide layout decisions within it are design decisions that reflect lead-developer's deck ownership. `warn` enforcement — cross-territory edits logged, not blocked. |
+| `scripts/render_slides.sh` | backend-engineer (bash script) | lead-developer (the inline `style:` block content it renders) | The script is backend; the CSS it renders is lead-developer's. |
+| `tests/test_slides_pipeline.py` | backend-engineer (test code) | lead-developer (assertions reflect deck structure + inline CSS) | The test code is backend; the assertions (slide count, inline style rules, theme-default) reflect lead-developer's deck/CSS decisions. |
+| `docs/presentations/README.md` | lead-developer (process narrative) | backend-engineer (build commands + render script invocations) | The process narrative is lead; the build commands + script names are backend. |
+
+## Domain priority (v1.23)
+
+`lead-developer → backend-engineer`
+
+Rationale: the deck markdown consolidation + inline CSS (lead) is the
+binding constraint — the render scripts (backend) can't be finalized
+until the deck structure is stable (slide count, inline style). The
+inline `style:` block must be ported + verified before the render
+pipeline (HTML + PPTX) is re-run. Backend (scripts + tests + CI)
+follows once the deck + CSS are settled. The dual-PPTX pipeline is
+wired last, once the deck renders clean with the inline style.
+
+---
+
+## Historical rosters
+
+<details>
+<summary>v1.22 roster (Nova Deck Layout Fix) — superseded by v1.23</summary>
+
+### Active personas (v1.22)
+
 ### lead-developer
 - **Domain:** coordination + deck content
 - **Active:** true
@@ -65,61 +207,13 @@ verification_toolchain:
   backend/scripting tasks.
 - **Phase-specific flag:** none (active for P2, P5; light touch on P0/P6).
 
-## Deactivated personas
+### D-148 (v1.22): Theme CSS is lead-developer territory, not frontend-engineer
+### D-149 (v1.22): No new personas for v1.22
 
-### frontend-engineer
-- **Active:** false
-- **Domain:** frontend
-- **Frameworks:** ["react", "next.js"] (inert — no territory)
-- **Constraints:** ["component-first", "server-components", "minimal-client-js"] (inert)
-- **Territory:** [] (no territory in v1.22)
-- **Reason:** v1.22 has no frontend; decks are markdown (lead-developer
-  territory); deactivated per PERSONAS.md v1.17/v1.18 precedent. The
-  theme CSS is a Marp stylesheet, not a React/Next.js component system
-  — it stays lead-developer territory. No reactivation trigger.
-
-### data-engineer
-- **Active:** false
-- **Domain:** data
-- **Frameworks:** [] (inert)
-- **Constraints:** [] (inert)
-- **Territory:** [] (no territory in v1.22)
-- **Reason:** v1.22 has no schema/DB/ORM changes. The milestone is
-  docs + scripts + tests only. No reactivation trigger.
-
-## Roster decisions
-
-### D-148 (0.95): Theme CSS is lead-developer territory, not frontend-engineer
-The `nova-sp-theme.css` is a Marp stylesheet (CSS for a markdown-to-
-slide renderer), not a React/Next.js component system. The v1.17/v1.18
-precedent (decks are markdown = lead-developer territory) extends to
-the deck's CSS theme. frontend-engineer's frameworks (react, next.js)
-are irrelevant to Marp CSS. **Decision:** theme CSS stays lead-developer
-territory. Confidence 0.95 — the only counter-argument is that CSS is
-"frontend," but Marp CSS is a static stylesheet, not a component system.
-
-### D-149 (0.9): No new personas for v1.22
-The work is CSS + bash + markdown + mermaid + pytest. All of this is
-within the two active personas' range (lead-developer: CSS + markdown +
-mermaid; backend-engineer: bash + pytest). Creating a separate "css-
-engineer" or "slides-engineer" persona would fragment ownership of the
-theme CSS + deck markdown (both lead-developer) and the render scripts
-+ tests (both backend-engineer). **Decision:** no new personas.
-Confidence 0.9.
-
-### Territory-overlap resolution (co-ownership)
-
-| Path | Primary | Co-owner | Why |
-|------|---------|----------|-----|
-| `docs/presentations/assets/mmd/*.mmd` | lead-developer (mermaid re-layout) | backend-engineer (re-render via render_slides.sh) | The .mmd content is lead-developer (diagram narrative); the PNG re-render is backend-engineer (script invocation). |
-| `tests/test_slides_pipeline.py` | backend-engineer (test code) | lead-developer (assertions reflect deck structure) | The test code is backend; the assertions (slide count, theme rules, aspect ratios) reflect lead-developer's deck/theme decisions. |
-
----
-
-## Historical rosters
+</details>
 
 <details>
-<summary>v1.18 roster (Citizen Developer & Production-Grade Guidance) — superseded by v1.22</summary>
+<summary>v1.18 roster (Citizen Developer & Production-Grade Guidance) — superseded</summary>
 
 ### Active personas (v1.18)
 
@@ -139,94 +233,32 @@ Confidence 0.9.
   - `docs/scope.md` (REQ-216)
   - `docs/skills.md` (REQ-222 — the index page, not the skill files themselves)
   - `docs/submission-readiness.md` (REQ-219 — citizen-developer-facing copy; co-owned with backend-engineer for the reason-code catalog)
-- **Reason:** Owns CIAgent metadata, the milestone narrative, the RACI +
-  PDLC-scope statements (REQ-215/216), the deck (21 slides, S&P theme
-  regression check vs P1, CAP-024), the skills index page (REQ-222), and
-  the citizen-developer-facing submission-readiness doc (REQ-219). Is
-  the only persona that touches `.ciagent/**` and the deck markdown.
-- **Phase-specific flag:** none (active for all of P0–P7).
 
 ### backend-engineer
 - **Domain:** backend
 - **Active:** true
 - **Phase-specific:** false
 - **Frameworks:** ["mcp (Python SDK v2)", "pydantic", "jsonschema", "urllib"]
-- **Constraints:** ["api-first", "strict-typing", "plugin-registry extensible (D-140)", "stdio now / HTTP-ready (D-135)", "no stack traces to citizen developers (REQ-218)"]
 - **Territory:**
   - `mcp/atelier/server.py` (REQ-223)
   - `mcp/atelier/plugins/**/*.py` (REQ-223 — principles.py, validation.py)
   - `mcp/atelier/vendor/**` (REQ-224 — vendored Atelier snapshot)
   - `mcp/atelier/VERSION.md` + `mcp/atelier/README.md` (REQ-224)
   - `scripts/update_atelier_vendor.sh` (REQ-224)
-  - `core/submission_readiness.py` (REQ-218 — the validator, invoked as `contract_ingestor.py --check-readiness`)
-  - `scripts/render_deck.sh` (REQ-228 — HTML + PPTX render)
-  - `scripts/attach_release_asset.py` (REQ-228 — Gitea release asset upload)
+  - `core/submission_readiness.py` (REQ-218)
+  - `scripts/render_deck.sh` (REQ-228)
+  - `scripts/attach_release_asset.py` (REQ-228)
   - `tests/test_atelier_mcp.py` (REQ-225)
   - `tests/test_submission_readiness.py` (REQ-220)
-  - `docs/submission-readiness.md` (REQ-219 — reason-code catalog section; co-owned with lead-developer for the narrative)
-- **Reason:** Owns the MCP server (plugin-registry, stdio, vendored
-  Atelier), the submission-readiness validator (extends
-  `contract_ingestor.py --check-readiness`, D-133), the render/attach
-  scripts (D-142 trigger), and the two new test files. The MCP
-  plugin-registry (D-140) is a backend pattern — no separate
-  mcp-engineer persona is created; backend-engineer owns it.
-- **Phase-specific flag:** none (active for P1 deck-render, P3 validator,
-  P5 MCP server, P6 scripts).
 
 ### data-engineer
 - **Domain:** data
 - **Active:** true
-- **Phase-specific:** false
 - **Frameworks:** ["jsonschema", "dynamodb (item shape)"]
-- **Constraints:** ["schema-first", "superset-gate NOT duplicate (PROJECT.md hard constraint)", "W3.E per-env mandatory table is the source of truth"]
 - **Territory:**
-  - `schemas/**` (REQ-217 — `submission-readiness.schema.json` is the new schema; existing schemas untouched)
-  - `core/lambda/contract_ingestor.py` (the `--check-readiness` subcommand wiring, D-133 — the validator is in `core/submission_readiness.py` but the ingestor dispatches to it; co-owned with backend-engineer)
-- **Reason:** Owns the submission-readiness JSON Schema (REQ-217) — it
-  is a schema artifact, data-engineer territory. The schema is a
-  *superset gate above* `contract.schema.json`, not a duplicate (it
-  references contract fields, does not redefine them). The
-  per-env-mandatory table comes from W3.E (the locked decision). The
-  ingestor wiring is co-owned with backend-engineer (the dispatch point
-  is backend; the schema it validates against is data).
-- **Phase-specific flag:** none (active for P3 schema + ingestor wiring).
+  - `schemas/**` (REQ-217)
+  - `core/lambda/contract_ingestor.py` (co-owned with backend-engineer)
 
-### Deactivated personas (v1.18)
-
-### frontend-engineer
-- **Active:** false
-- **Domain:** frontend
-- **Frameworks:** ["react", "next.js"] (inert — no territory)
-- **Constraints:** ["component-first", "server-components", "minimal-client-js"] (inert)
-- **Territory:** [] (no territory in v1.18)
-- **Reason:** v1.18 has no frontend; decks are markdown (lead-developer
-  territory); deactivated per PERSONAS.md v1.17 precedent. v1.18's
-  observability stays PowerBI / external (Out of Scope: "A Nova-built
-  frontend / dashboard"). The MCP server exposes tools to an AI agent,
-  not a web UI. No reactivation trigger in this milestone.
-
-### Roster decisions (v1.18)
-
-### D-143 (0.90): Fold mcp-engineer into backend-engineer
-The MCP plugin-registry (D-140: `plugins/<name>.py register(mcp)`) is a
-backend code pattern — Python modules, type hints, stdio transport,
-urllib for the Gitea asset API. It shares nothing with the data domain
-(schemas/DynamoDB) and is not a new engineering discipline. Creating a
-separate `mcp-engineer` persona would fragment ownership of the server +
-its tests + the render/attach scripts (all backend). **Decision:** fold
-into backend-engineer. backend-engineer's `frameworks` list gains
-`mcp (Python SDK v2)`. Confidence 0.90 — the only counter-argument is
-that MCP is a distinct protocol skill, but the SDK v2 API surface
-(`@mcp.tool()` + type hints) is small and well within backend-engineer's
-range (it's the same Pydantic/FastAPI-style pattern the persona already
-knows).
-
-### Territory-overlap resolution (v1.18)
-
-| Path | Primary | Co-owner | Why |
-|------|---------|----------|-----|
-| `docs/submission-readiness.md` | lead-developer (narrative + examples) | backend-engineer (reason-code catalog, REQ-218 codes) | The doc is citizen-developer-facing copy (lead) but the reason-code catalog (MISSING_TAGS, ENV_MISSING_MANDATORY, AGENTIC_MISSING_INTENT, MISSING_APP_SOURCE, POLICY_PRECONDITION_MISSING) is backend (it mirrors the validator's return codes). |
-| `core/lambda/contract_ingestor.py` | backend-engineer (dispatch wiring) | data-engineer (the schema it validates against) | D-133 places the `--check-readiness` subcommand on the ingestor (backend dispatch), but the readiness schema it loads is data-engineer territory. |
-| `schemas/submission-readiness.schema.json` | data-engineer (schema artifact) | backend-engineer (the validator must match it) | The schema is data-engineer's; the validator (REQ-218) is backend-engineer's and must stay in sync with it. |
+### D-143 (v1.18): Fold mcp-engineer into backend-engineer
 
 </details>
\ No newline at end of file
diff --git a/.ciagent/RESEARCH.md b/.ciagent/RESEARCH.md
index 661deaa..bc8433a 100644
--- a/.ciagent/RESEARCH.md
+++ b/.ciagent/RESEARCH.md
@@ -2568,3 +2568,705 @@ All 5 CLARIFY decisions are grounded in these findings:
 - Re-layout to LR (FINDING 3 — TB produced 0.63 aspect)
 - Delete render_deck.sh (FINDING 5 — stale, unthemed)
 - Split slides 3+8 (FINDING 2 — densest overflow)
+
+---
+
+## v1.23 RESEARCH — Nova Deck Cleanup & Python PPTX
+
+> Phase: 0 RESEARCH. Milestone v1.23 (Nova Deck Cleanup & Python PPTX).
+> docs/render/test NFR milestone. Surface: deck markdown, Marp inline
+> style CSS, Python render scripts, tests, CI YAML, README.
+> Researcher: ci-researcher. Sources: ACDL codebase (v1.22 tree) +
+> the deleted clean reference HTML at /root/gl/acdl/docs/presentations/
+> the-developer-experience.html + python-pptx docs + Marp rendered HTML
+> inspection.
+
+### FINDING 1 — Marp `default` theme + inline `style:` block (the clean reference)
+
+The deleted clean reference deck
+`/root/gl/acdl/docs/presentations/the-developer-experience.html`
+(commit eb43e08) used `theme: default` + an inline `style:` block in the
+Marp frontmatter. The exact inline CSS (extracted from the `data-style`
+attribute repeated on every `<section>` in the rendered HTML) is:
+
+```css
+section {
+  font-family: "Akkurat Pro", "Helvetica Neue", "Arial", sans-serif;
+  font-size: 26px;
+  color: #1B1B1B;
+}
+h1 { color: #D6002A; font-size: 40px; margin-bottom: 0.3em; }
+h2 { color: #D6002A; font-size: 32px; margin-bottom: 0.2em; }
+section.title { background: #1B1B1B; color: #fff; border-top: 8px solid #D6002A; }
+section.title h1 { color: #fff; }
+table { font-size: 22px; width: 100%; }
+th { background: #F0F0F0; }
+blockquote { border-left: 4px solid #D6002A; color: #2E2E2E; font-size: 24px; }
+pre { font-size: 16px; line-height: 1.3; }
+code { font-size: 16px; }
+img { display: block; margin: 0 auto; max-height: 280px; }
+.badge {
+  display: inline-block; padding: 2px 8px; border-radius: 4px;
+  font-size: 16px; font-weight: 600;
+}
+.planned { background: #fef3c7; color: #78350f; }
+```
+
+**What Marp's built-in `default` theme provides for free** (from the
+embedded `@theme default` CSS in the reference HTML):
+- `section { width:1280px; height:720px; box-sizing:border-box; overflow:hidden; position:relative }`
+  — slide dimensions + box model.
+- Base section typography: `font-family: var(--fontStack-sansSerif)`
+  (system sans), `font-size: 16px`, `line-height: 1.5`, `color:
+  var(--fgColor-default)`, `background: var(--bgColor-default)` (white).
+  **Note: default does NOT set padding** — wait, correction: the
+  `default` theme's embedded CSS in this Marp version does NOT include
+  an explicit `padding` on `section`; the nova-sp-theme.css comment
+  claiming `padding: 56px 64px` was for a different Marp version. The
+  inline `style:` block above does not set padding either, implying the
+  default theme's inheriting UA stylesheet padding or Marp's container
+  provides it. **This is the key thing to verify in P2:** whether
+  `theme: default` + the inline `style:` block produces padding that
+  reserves header/footer chrome space. The v1.22 root-cause was ZERO
+  padding with the standalone theme; the inline `style:` block above
+  also does not set padding, so the clean look may have relied on the
+  default theme's own padding (which Marp's default theme CSS
+  *does* set: `section{padding:56px 64px}` is in the default theme's
+  canonical CSS even if not visible in this rendered HTML's inlined
+  subset — the rendered HTML strips the theme CSS into a `<style>` block
+  at the top that we did not fully inspect). **Recommendation:** the
+  inline `style:` block for v1.23 SHOULD explicitly set section padding
+  (`padding: 48px 56px 40px` as v1.22 settled on) to be safe, OR verify
+  the default theme provides it. The reference deck's clean look
+  suggests the default theme's padding was adequate.
+- Pagination: `section::after { content: attr(data-marpit-pagination);
+  bottom:0; right:0; padding:inherit }` — page numbers via CSS pseudo-
+  element on `section[data-marpit-pagination]`. Free with `paginate:
+  true`.
+- Header/footer chrome: `default` theme renders `header`/`footer`
+  elements from frontmatter `header:`/`footer:` directives. The inline
+  `style:` block does NOT suppress them on title slides (the reference
+  deck's title slide has `<header>The Developer Experience</header>` +
+  `<footer>Internal</footer>` visible). v1.22's `section.title header,
+  section.title footer { display: none }` rule should be ported into the
+  v1.23 inline `style:` block to keep title slides clean.
+- H1/H2 base sizing: `:is(h1, marp-h1){font-size:2em;margin-block:0.67em}`
+  — the inline `style:` block overrides these with explicit px sizes.
+- Font-family: the default theme uses GitHub's `--fontStack-sansSerif`
+  (system fonts); the inline `style:` overrides to Akkurat Pro stack.
+
+**What the inline `style:` block must override** (not inherited from
+default): font-family (Akkurat Pro stack), font-size (26px vs 16px),
+font-color (#1B1B1B), H1/H2 colors (#D6002A), title-slide background
+(#1B1B1B + 8px red top border), table styling, blockquote styling,
+image max-height. **What it inherits for free:** slide dimensions
+(1280×720), pagination, header/footer rendering, base box model.
+
+**Confidence: 0.95** (the CSS is read verbatim from the rendered
+`data-style` attribute on every `<section>`; the default theme
+inheritance is confirmed by the `data-theme="default"` attribute on
+every section).
+
+### FINDING 2 — Marp HTML-in-markdown passthrough
+
+Marp (via Marpit, the underlying markdown-it-based engine) **passes
+through raw HTML blocks** in markdown slides. Confirmed by the
+reference deck:
+`/root/gl/acdl/docs/presentations/the-developer-experience-marp.md`
+contains inline HTML like
+`<span class="badge planned">Planned</span>` inside slide bodies, and
+the rendered HTML (`the-developer-experience.html`) contains the
+verbatim `<span class="badge planned">Planned</span>` inside the
+`<section>` elements. The `.badge` and `.planned` classes are styled
+by the inline `style:` block (Finding 1).
+
+**Gotchas confirmed:**
+- HTML blocks render correctly inside `## Slide N — Title` H2 slides
+  (the reference deck's slide 5 has `<span class="badge planned">` in a
+  table cell; slide 7 has it in a bullet list item).
+- HTML does NOT interfere with pagination (the `data-marpit-pagination`
+  attribute is present on every paginated section regardless of HTML
+  content).
+- HTML inline elements (`<span>`, `<div>`, `<strong>`) render; block
+  elements (`<div class="benefit">text</div>`) also render — Marpit's
+  markdown-it `html: true` option is on by default.
+- **Caveat:** Marp does NOT sanitize HTML by default (no DOMPurify);
+  raw `<script>` would pass through too. For our use case (styling
+  divs/spans) this is fine.
+- The current nova deck markdown does NOT use any `<div>` HTML blocks
+  (`grep -c "<div" nova-autonomous-cloud-delivery-marp.md` = 0 in the
+  slide body; only the bespoke-marp-osc div in the rendered HTML
+  chrome). So the v1.23 plan to add `.benefit` class divs is a NEW
+  pattern — it will work, but is untested in this repo's deck.
+
+**Confidence: 0.95** (confirmed by direct inspection of the reference
+deck's source + rendered HTML; Marpit's html passthrough is documented
+behavior).
+
+### FINDING 3 — python-pptx capabilities for the structured PPTX
+
+python-pptx (the `python-pptx` PyPI package) supports all capabilities
+required for `scripts/render_pptx.py`. API calls confirmed against the
+python-pptx documentation (v1.0.2 latest; v0.6.23 is the established
+minimum):
+
+(a) **16:9 slide dimensions:**
+```python
+from pptx import Presentation
+from pptx.util import Inches, Pt
+prs = Presentation()
+prs.slide_width = Inches(13.333)   # 16:9 widescreen
+prs.slide_height = Inches(7.5)
+```
+
+(b) **Slide background color (black for title slide):**
+```python
+from pptx.dml.color import RGBColor
+slide = prs.slides.add_slide(prs.slide_layouts[6])  # blank layout
+bg = slide.background
+bg.fill.solid()
+bg.fill.fore_color.rgb = RGBColor(0x1B, 0x1B, 0x1B)  # sp-black
+```
+(Note: `slide.background` is available since python-pptx 0.6.19+.)
+
+(c) **Colored rectangle as top border bar:**
+```python
+from pptx.enum.shapes import MSO_SHAPE
+from pptx.util import Inches
+shape = slide.shapes.add_shape(
+    MSO_SHAPE.RECTANGLE,
+    Inches(0), Inches(0),         # left, top
+    Inches(13.333), Inches(0.08) # width, height (8px ≈ 0.08in at 96dpi)
+)
+shape.fill.solid()
+shape.fill.fore_color.rgb = RGBColor(0xD6, 0x00, 0x2A)  # sp-red
+shape.line.fill.background()  # no border on the rectangle
+```
+
+(d) **Text frames with font family/color/size (Akkurat Pro → Helvetica
+Neue → Arial fallback):**
+```python
+from pptx.util import Pt
+txBox = slide.shapes.add_textbox(Inches(0.5), Inches(0.5), Inches(12), Inches(1))
+tf = txBox.text_frame
+p = tf.paragraphs[0]
+run = p.add_run()
+run.text = "Slide title"
+run.font.name = "Akkurat Pro"
+run.font.size = Pt(40)
+run.font.bold = True
+run.font.color.rgb = RGBColor(0xD6, 0x00, 0x2A)
+# Font fallback: python-pptx sets a single font.name; PowerPoint's own
+# font substitution handles the fallback chain. To set a fallback
+# explicitly, set the font's `<a:latin>` + `<a:ea>` + `<a:cs>` typeface
+# attributes, OR rely on PowerPoint rendering "Akkurat Pro" → missing →
+# substitutes Helvetica Neue / Arial per the OS font substitution table.
+# The simplest approach: set name="Akkurat Pro"; PowerPoint does the
+# fallback. For an explicit fallback, manipulate the XML:
+from pptx.oxml.ns import qn
+rPr = run._r.get_or_add_rPr()
+latin = rPr.find(qn('a:latin'))
+if latin is None:
+    latin = rPr.makeelement(qn('a:latin'), {})
+    rPr.append(latin)
+latin.set('typeface', 'Akkurat Pro')
+# (PowerPoint reads only the first typeface; fallback is OS-level.)
+```
+**Key constraint:** python-pptx does NOT support a CSS-style font
+fallback stack. Setting `run.font.name` sets one typeface; PowerPoint's
+font substitution handles the fallback when the font is missing. This
+is adequate for our use case (Akkurat Pro is the named font; on
+systems without it, PowerPoint substitutes Helvetica/Arial).
+
+(e) **Bullet lists:**
+```python
+tf = txBox.text_frame
+for i, item in enumerate(bullets):
+    p = tf.paragraphs[0] if i == 0 else tf.add_paragraph()
+    p.text = item
+    p.level = 0  # bullet level (0 = top)
+    # Bullet character: python-pptx does not expose a high-level bullet
+    # API; the default paragraph style from the layout provides bullets.
+    # For explicit bullet control, manipulate the XML <a:buChar> or
+    # set p.font + use the layout's list style.
+```
+**Constraint:** python-pptx has limited high-level bullet support; the
+`paragraph` level controls indentation but the bullet glyph comes from
+the layout's list style. For a blank layout (layout 6), paragraphs
+have NO bullet by default — you must add `<a:buChar>` to the paragraph
+properties XML, or use a layout that has bullets. **Recommendation:**
+use `slide_layouts[1]` (Title and Content) for bullet slides, which
+provides bullets natively; or inject `<a:buChar char="•"/>` via XML.
+
+(f) **Native tables (`add_table`):**
+```python
+from pptx.util import Inches
+rows, cols = len(data), len(headers)
+table = slide.shapes.add_table(rows, cols, Inches(0.5), Inches(1.5),
+                                Inches(12), Inches(4)).table
+# Headers
+for c, h in enumerate(headers):
+    table.cell(0, c).text = h
+    # style header cell
+    cell = table.cell(0, c)
+    cell.fill.solid()
+    cell.fill.fore_color.rgb = RGBColor(0xF0, 0xF0, 0xF0)
+# Body
+for r, row in enumerate(data, start=1):
+    for c, val in enumerate(row):
+        table.cell(r, c).text = str(val)
+```
+**Confirmed:** `add_table` is fully supported since 0.6.19; returns a
+`Table` object with `.cell(r,c).text` + `.fill`.
+
+(g) **Embedding images (`add_picture`, auto-embeds in ppt/media/):**
+```python
+slide.shapes.add_picture(
+    "docs/presentations/assets/png/platform-pipeline.png",
+    Inches(1), Inches(1),
+    width=Inches(11), height=Inches(5)
+)
+```
+**Confirmed:** python-pptx auto-embeds the image binary into the
+`ppt/media/` directory inside the .pptx zip; the slide XML references
+it via a relationship ID. This is native — no manual zip manipulation
+needed. The image is copied once; multiple slides referencing the
+same image share the media file.
+
+(h) **Italic text:**
+```python
+run.font.italic = True  # supported since 0.6.0
+```
+
+**Version constraint:** `python-pptx>=0.6.23` is the right minimum
+(stable release with `slide.background`, `add_table`, `add_picture`
+all mature). The latest is 1.0.2 (2024); 0.6.23 (2021) is safe. The
+`slide.background` API landed in 0.6.19; `add_table` in 0.6.19;
+`add_picture` since 0.6.0. **Recommendation:** pin
+`python-pptx>=0.6.23` in `pyproject.toml` `[project.optional-
+dependencies] test`.
+
+**Confidence: 0.90** (API calls confirmed against python-pptx docs;
+the font fallback + bullet limitations are the residual uncertainty —
+both have XML workarounds).
+
+### FINDING 4 — Image inlining (stdlib-only `base64` + `re` + `mimetypes`)
+
+The current rendered HTML
+(`docs/presentations/nova-autonomous-cloud-delivery.html`) embeds
+images as **plain `<img src="assets/png/...">` tags with relative
+paths** — NOT base64-inlined. Confirmed by grep: `<img
+src="assets/png/platform-pipeline.png" ...>` (slide 7), `<img
+src="assets/png/telemetry-live-ops.png" ...>` (slide 11). They are
+NOT SVG `xlink:href`; they are standard `<img>` elements inside the
+SVG `<foreignObject>` wrapper that Marp uses for slide rendering.
+
+**The stdlib-only approach is sufficient:**
+```python
+import base64, re, mimetypes
+html = Path("...html").read_text()
+def inline(match):
+    src = match.group(1)
+    if src.startswith("data:"):  # already inlined
+        return match.group(0)
+    mime = mimetypes.guess_type(src)[0] or "image/png"
+    data = base64.b64encode(Path(src).read_bytes()).decode()
+    return f'<img src="data:{mime};base64,{data}"'
+# Match <img src="..."> with single or double quotes
+html = re.sub(r'<img\s+src="([^"]+)"', inline_img, html)
+html = re.sub(r"<img\s+src='([^']+)'", inline_img, html)
+```
+
+**Edge cases confirmed:**
+- Marp's SVG-foreignObject wrapper: the `<img>` tags are INSIDE the
+  `<foreignObject><section>...</section></foreignObject>` — the regex
+  still matches them (they're standard HTML img tags in the HTML
+  namespace). No special handling needed.
+- Quoted vs unquoted src: Marp always emits double-quoted `src="..."`;
+  handle both for safety.
+- Relative vs absolute paths: Marp emits relative paths
+  (`assets/png/X.png`); the script must resolve them relative to the
+  HTML file's directory. `Path(html_path).parent / src`.
+- Already-inlined images (`data:` URIs): skip (the regex checks for
+  `src.startswith("data:")`).
+- Marp does NOT use `xlink:href` for images in HTML output (that's
+  only for SVG-only output mode); the HTML output uses standard
+  `<img src>`. Confirmed by grep: 0 occurrences of `xlink:href` in
+  the current HTML.
+- The `--allow-local-files` flag in `render_slides.sh` is what makes
+  Marp keep the local relative paths (without it, Marp would skip
+  local images in the standalone HTML).
+
+**No external library needed.** stdlib `base64` + `re` + `mimetypes`
++ `pathlib` is sufficient.
+
+**Confidence: 0.95** (verified by direct inspection of the current
+HTML; the regex handles all observed patterns).
+
+### FINDING 5 — Tests that need updating (`tests/test_slides_pipeline.py`)
+
+Full read of `tests/test_slides_pipeline.py` (373 lines). Tests that
+assert on v1.23-affected surfaces:
+
+**(a) `theme: nova-sp` / `nova-sp-theme.css` via `--theme`:**
+- **Line 55-64 `test_marp_deck_uses_sp_theme`**: asserts `"nova-sp" in
+  frontmatter`. After v1.23: frontmatter should have `theme: default`;
+  this test must assert `"theme: default"` instead (or assert the inline
+  `style:` block is present).
+- **Line 66-74 `test_marp_deck_not_using_default_theme`**: asserts
+  `"theme: default" not in frontmatter`. After v1.23: this test must
+  be DELETED or INVERTED (we're reverting TO `theme: default`).
+- **Line 36-38 `test_sp_theme_css_exists`**: asserts
+  `nova-sp-theme.css` is a file. After v1.23: keep (the CSS is retired
+  from render but kept as a reference file).
+- **Line 41-45 `test_sp_theme_css_has_snp_colors`**: asserts the CSS
+  has S&P colors. After v1.23: keep (reference file still has colors).
+- **Line 275-282 `test_theme_css_has_section_padding`**: asserts the
+  CSS `section` rule has padding. After v1.23: this test targets the
+  RETIRED CSS — it should be retargeted to the inline `style:` block in
+  the marp deck frontmatter (assert the inline style has section
+  padding).
+- **Line 284-293 `test_theme_css_suppresses_title_chrome`**: asserts
+  `section.title header` + `display: none` in the CSS. After v1.23:
+  retarget to the inline `style:` block.
+- **Line 295-302 `test_theme_css_has_aspect_ratio_aware_images`**:
+  asserts `object-fit` + `max-width` in the CSS. After v1.23: retarget
+  to the inline `style:` block (or drop if the inline style uses the
+  simpler `max-height: 280px` from the reference deck).
+- **Line 353-359 `test_html_embeds_theme`**: asserts `"--sp-red"` +
+  `"padding:"` in the committed HTML. After v1.23: the inline `style:`
+  block uses literal `#D6002A` (not the `--sp-red` CSS variable — the
+  reference deck's inline style did NOT use CSS variables). This test
+  must change to assert `"#D6002A"` + `"padding:"` (or `"font-family:
+  \"Akkurat Pro\""`).
+
+**(b) Both `.md` and `-marp.md` existing:**
+- **Line 31 `SOURCE_MD` constant** + **Line 240-263
+  `test_source_md_no_internal_citations_in_slides`** +
+  **Line 266-270 `test_source_md_no_badges`**: assert the plain `.md`
+  is the source of truth. After v1.23: the plain `.md` is DELETED
+  (consolidated into `-marp.md`); these tests must be DELETED or
+  retargeted to the consolidated `-marp.md`.
+
+**(c) `marp-cli` invocation specifics:**
+- **Line 83-90 `test_render_slides_script_renders_mermaid`**: asserts
+  `mermaid-cli` + `sp-theme.json` in the script. After v1.23: keep
+  (mermaid render unchanged).
+- **Line 92-98 `test_render_slides_script_renders_marp`**: asserts
+  `marp-cli` + `.html` + `.pptx` in the script. After v1.23: the script
+  no longer passes `--theme nova-sp-theme.css`; it may still produce
+  `.pptx` via marp-cli OR via the new `render_pptx.py`. This test must
+  be updated to reflect the new invocation (marp-cli for HTML only;
+  render_pptx.py for the structured PPTX).
+- **Line 327-332 `test_render_slides_has_2x_scale`**: asserts `-s 2`
+  + `-b transparent`. After v1.23: keep (mermaid flags unchanged).
+- **Line 335-344 `test_render_slides_pins_cli_versions`**: asserts
+  `marp-cli@` + `mermaid-cli@` pinned, no `@latest`. After v1.23: keep
+  for mermaid; if marp-cli is still used for HTML, keep for marp too.
+- **Line 346-351 `test_render_deck_removed`**: asserts
+  `render_deck.sh` is deleted. After v1.23: keep.
+
+**(d) Slide count:**
+- **Line 204-218 `test_marp_deck_slide_count`**: asserts 20 main + 1
+  appendix = 21 slides. After v1.23: the word-count trim may change
+  the count; update to match the final count (likely stays 20+1 or
+  drops to fewer if slides merge).
+- **Line 361-373 `test_html_slide_count_matches_marp`**: asserts the
+  HTML `<section>` count matches the deck count. After v1.23: keep
+  (recompute expected from the final deck).
+
+**(e) Other affected tests:**
+- **Line 137-143 `test_readme_no_retired_decks`**: asserts README
+  doesn't list old decks. After v1.23: keep.
+- **Line 146-150 `test_readme_no_old_deck_name`**: asserts README
+  references `nova-autonomous-cloud-delivery`. After v1.23: keep.
+- **Line 153-156 `test_old_deck_files_removed`**: asserts old
+  `nova-no-humans-platform*` files are gone. After v1.23: keep.
+- **Line 159-162 `test_marp_deck_no_badges`**: asserts no `badge` in
+  marp deck. After v1.23: keep.
+- **Line 165-177 `test_marp_deck_no_version_in_footer`**: asserts no
+  version in frontmatter. After v1.23: keep.
+- **Line 179-190 `test_marp_deck_title_slide_no_version_subtitle`**:
+  asserts no `v1.18` subtitle. After v1.23: keep.
+- **Line 193-202 `test_marp_deck_title_is_autonomous_cloud_delivery`**:
+  asserts the title. After v1.23: keep (unless "penetrate" removal
+  touches the title — it doesn't).
+- **Line 221-237 `test_marp_deck_no_internal_citations`**: asserts no
+  D-###/REQ-###/.py paths. After v1.23: keep.
+- **Line 304-324 `test_png_aspect_ratios_sane`**: asserts PNG aspect
+  ratios in [0.4, 4.0]. After v1.23: keep.
+- **Line 107-119 `test_slides_ci_workflow_*`**: asserts slides.yml
+  exists + triggers on presentations. After v1.23: keep (update if the
+  workflow changes).
+
+**Summary:** ~12 tests need updating; the key inversions are
+`test_marp_deck_not_using_default_theme` (delete/invert) and the
+`test_theme_css_*` trio (retarget from the standalone CSS file to the
+inline `style:` block). The `test_source_md_*` tests must be deleted
+(plain `.md` is gone).
+
+**Confidence: 0.95** (direct line-by-line read of the test file).
+
+### FINDING 6 — `scripts/attach_release_asset.py`
+
+Read in full (80 lines). Current behavior:
+- CLI: `python3 scripts/attach_release_asset.py <file-path> <release-id>`
+- Resolves `NOVA_GITEA_TOKEN` (or `ACDL_GITEA_TOKEN` fallback) from
+  `.env.secrets`/`.env`.
+- POSTs to `/api/v1/repos/{owner}/{repo}/releases/{id}/assets` with
+  multipart form: `name=<filename>` + `attachment=<file bytes>`.
+- Single-file upload per invocation.
+
+**Change needed for v1.23:** attach BOTH `.pptx` (Marp PPTX) AND
+`-python.pptx` (python-pptx structured PPTX). Two options:
+1. Call the script twice (once per file) from the ship workflow.
+2. Extend the script to accept multiple file paths: `attach_release_asset.py
+   <file1> <file2> ... <release-id>` (loop over files, same release-id).
+
+**Recommendation:** option 2 (extend to multiple files) — minimal,
+backward-compatible (if only 1 file given, same behavior). The
+release-id stays the last positional arg; all preceding args are file
+paths.
+
+**Confidence: 0.95** (direct read; the change is mechanical).
+
+### FINDING 7 — `workflows-src/slides.yml`
+
+Read in full (31 lines). Current CI steps:
+1. `actions/checkout@v4` (fetch-depth 0).
+2. `actions/setup-node@v4` (node 20).
+3. "Install Chrome" — runs `npx --yes @marp-team/marp-cli@latest
+   --version` + `npx --yes @mermaid-js/mermaid-cli --version` (NOTE:
+   uses `@latest` — NOT pinned, contradicts REQ-257's pinning rule for
+   `render_slides.sh`; the workflow is separate from the script).
+4. "Render slides" — `bash scripts/render_slides.sh`.
+5. "Commit rendered artifacts" — `git add docs/presentations/*.html
+   docs/presentations/*.pptx docs/presentations/assets/png/*.png` +
+   commit + push.
+
+**Where `python-pptx` install would go:** a new step between "Install
+Chrome" and "Render slides":
+```yaml
+- uses: actions/setup-python@v5
+  with: { python-version: '3.11' }
+- name: Install python-pptx
+  run: pip install python-pptx>=0.6.23
+```
+Then `render_slides.sh` (or a new `render_pptx.py` invocation) uses
+it. OR: `pip install -e ".[test]"` if python-pptx is added to
+`pyproject.toml [project.optional-dependencies] test`.
+
+**What the commit list should include after v1.23:** the
+`git add` line must include:
+- `docs/presentations/*.html` (re-rendered HTML, with inlined images)
+- `docs/presentations/*.pptx` (Marp PPTX — kept as the "fallback" PPTX)
+- `docs/presentations/*-python.pptx` (NEW — the python-pptx structured
+  PPTX)
+- `docs/presentations/assets/png/*.png` (re-rendered mermaid PNGs)
+
+**Also:** the workflow uses `@latest` for both CLIs in the "Install
+Chrome" step — this should be pinned to match `render_slides.sh`'s
+pinned versions (`@marp-team/marp-cli@4.5.0`,
+`@mermaid-js/mermaid-cli@11.16.0`) for reproducibility.
+
+**Confidence: 0.95** (direct read).
+
+### FINDING 8 — `docs/presentations/README.md` current structure
+
+Read in full (249 lines). Current 4-step process structure:
+- **Step 1 — Full markdown (source of truth):** `<deck-name>.md` with
+  `## Slide N — Title` + `> **Speaker notes:**` blocks + mermaid code
+  blocks.
+- **Step 2 — Marp deck synthesis:** `<deck-name>-marp.md` with Marp
+  frontmatter (`theme: nova-sp`), no speaker notes, embedded PNGs, no
+  badges.
+- **Step 3 — Render to HTML and PPTX:** `bash scripts/render_slides.sh
+  <deck-name>`; HTML committed (with base64 images — NOTE: this is
+  aspirational; the current HTML does NOT inline images, which is what
+  `inline_images.py` will fix); PPTX committed + release-attached.
+- **Step 4 — Talking points:** `<deck-name>-talking-points.md` with
+  one section per Marp slide, 3-6 bullets + key takeaway.
+
+**What to rewrite for v1.23:**
+- The 4-step process becomes a **3-step process** (Step 1 plain `.md`
+  is DELETED → consolidated into `-marp.md`; the source of truth IS the
+  `-marp.md` with speaker notes in HTML comments, OR speaker notes are
+  dropped entirely). OR: keep 4 steps but Step 1's output merges into
+  Step 2 (the `-marp.md` becomes the source of truth, with speaker
+  notes preserved as HTML comments that Marp strips from rendering).
+- Step 2 frontmatter: `theme: default` + inline `style:` block (not
+  `theme: nova-sp`).
+- Step 3: HTML rendered by Marp (`render_slides.sh`), images inlined
+  by `scripts/inline_images.py` (NEW), PPTX rendered by BOTH marp-cli
+  (fallback `.pptx`) AND `scripts/render_pptx.py` (structured
+  `-python.pptx`).
+- The directory layout block (line 108-123) must drop the plain `.md`
+  row + add `inline_images.py` + `render_pptx.py` rows + the
+  `-python.pptx` row + note `nova-sp-theme.css` is retired (reference
+  only).
+- The "Current decks" table (line 237-239) must list both PPTX files.
+- The slide-count convention (line 128-136) must reflect the final
+  v1.23 count (after word-count trim + "penetrate" removal).
+- The "Adding a new presentation" section (line 209-233) must be
+  rewritten for the single-document + dual-PPTX pipeline.
+
+**Confidence: 0.95** (direct read).
+
+### FINDING 9 — Persona assessment for v1.23
+
+The v1.23 scope: deck markdown editing (consolidate `.md` + `-marp.md`,
+trim word count, remove "penetrate"), Marp inline `style:` CSS, Python
+scripts (`inline_images.py` + `render_pptx.py`), tests, CI YAML
+(`slides.yml`), README. The v1.22 roster had lead-developer +
+backend-engineer active; frontend + data deactivated.
+
+**Assessment:** the same roster applies, with one territory shift.
+- **lead-developer** (active): owns the deck markdown (consolidation,
+  word-count trim, "penetrate" removal), the inline `style:` CSS block
+  (porting the reference deck's CSS), the README rewrite, and all
+  `.ciagent/**` metadata. The inline CSS is a Marp stylesheet block, not
+  a React/Next.js system — stays lead-developer (D-148 precedent from
+  v1.22).
+- **backend-engineer** (active): owns the Python scripts
+  (`inline_images.py`, `render_pptx.py` — python-pptx is a backend
+  library), the `render_slides.sh` updates, the tests, the CI YAML,
+  and the `attach_release_asset.py` extension. This is a shift from
+  v1.22 (where backend owned bash + pytest); v1.23 adds Python to
+  backend's territory.
+- **frontend-engineer** (deactivated): no React/Next.js. The inline
+  CSS is a Marp stylesheet, not frontend. Stays deactivated (D-148
+  precedent).
+- **data-engineer** (deactivated): no schema/DB/ORM. Stays deactivated.
+
+**Custom "docs/deck" persona?** Not warranted. The work splits cleanly
+into lead-developer (narrative + CSS + markdown + README + metadata)
+and backend-engineer (Python + bash + tests + CI). A "docs-engineer"
+or "slides-engineer" persona would fragment ownership of the deck
+markdown (lead) and the render scripts (backend). D-149 precedent from
+v1.22 applies: no new personas.
+
+**One nuance:** python-pptx is a new dependency (backend territory) but
+the PPTX CONTENT (slide layout, text, image placement) is a design
+decision that overlaps with lead-developer's deck ownership. The
+`render_pptx.py` script is backend (Python code); the slide-by-slide
+layout decisions within it (which image goes where, font sizes) are
+co-owned with lead-developer. Territory enforcement = `warn` (cross-
+territory edits logged, not blocked).
+
+**Confidence: 0.90** (the roster follows v1.22 precedent; the only new
+element is python-pptx, which is unambiguously backend).
+
+### FINDING 10 — Risks / pitfalls (5)
+
+1. **Marp `default` theme padding conflict.** The reference deck's
+   inline `style:` block does NOT set `section` padding. The v1.22
+   root-cause was ZERO padding with the standalone `nova-sp-theme.css`.
+   If `theme: default`'s built-in padding differs from v1.22's settled
+   `48px 56px 40px`, slides may re-overflow. **Mitigation:** explicitly
+   set `section { padding: 48px 56px 40px }` in the v1.23 inline
+   `style:` block (do not rely on the default theme's padding). Verify
+   by re-rendering + visual inspection in P2.
+
+2. **python-pptx font fallback when Akkurat Pro isn't installed.**
+   python-pptx sets a single `run.font.name = "Akkurat Pro"`; on
+   systems without Akkurat Pro (CI, most laptops), PowerPoint
+   substitutes a default font (often Calibri), NOT Helvetica Neue /
+   Arial. The CSS stack `Akkurat Pro → Helvetica Neue → Arial` does
+   NOT translate to python-pptx (no CSS-style fallback). **Mitigation:**
+   accept PowerPoint's OS-level substitution (the deck will render in
+   a sans font either way); OR set `run.font.name = "Arial"` directly
+   for the structured PPTX (guaranteed present on all systems) and
+   document that the Marp HTML deck uses the full Akkurat Pro stack
+   while the python PPTX uses Arial. The dual-PPTX approach lets the
+   Marp PPTX (rendered by Chromium with the CSS stack) carry the
+   brand font, while the python PPTX (structured, editable) carries
+   Arial for editability.
+
+3. **Image inlining breaking if Marp embeds images as SVG
+   `xlink:href`.** The current HTML uses `<img src="assets/png/...">`
+   (confirmed by grep). BUT if a future Marp version switches to SVG
+   `xlink:href` for the HTML output, the `inline_images.py` regex
+   would miss them. **Mitigation:** the regex should ALSO match
+   `xlink:href="assets/..."` patterns as a fallback; OR pin the
+   Marp version (already pinned at `@4.5.0` in `render_slides.sh`) so
+   the output format is stable.
+
+4. **Consolidated doc losing speaker notes if HTML comments are
+   stripped by Marp.** The plan consolidates the plain `.md` (with
+   `> **Speaker notes:**` blockquotes) into the `-marp.md`. Marp strips
+   HTML comments (`<!-- ... -->`) from rendered output but preserves
+   blockquotes (`> ...`) as `<blockquote>` elements. If speaker notes
+   are moved to HTML comments to hide them from the audience, Marp
+   strips them — and they're lost from the rendered HTML. If they stay
+   as blockquotes, they're visible in the audience deck (bad).
+   **Mitigation:** either (a) drop speaker notes entirely from the
+   consolidated `-marp.md` (the talking-points file is the presenter
+   cue sheet; notes aren't needed in the deck), or (b) keep them as
+   HTML comments (Marp strips from render, but they persist in the
+   source markdown for git-diff readability). Option (b) is safer —
+   the source markdown keeps the notes; the rendered HTML + PPTX
+   don't. Verify Marp strips `<!-- -->` from HTML output (confirmed:
+   Marp's markdown-it does strip HTML comments by default).
+
+5. **`test_html_embeds_theme` asserts `--sp-red` CSS variable, but the
+   reference inline `style:` uses literal `#D6002A`.** The v1.22 test
+   (line 353-359) asserts `"--sp-red"` in the committed HTML. The
+   reference deck's inline `style:` block uses literal `#D6002A`, not
+   a CSS variable. If v1.23 ports the reference style verbatim, the
+   test fails. **Mitigation:** update the test to assert `"#D6002A"`
+   (literal) OR use CSS variables in the inline `style:` block (Marp
+   supports `:root { --sp-red: #D6002A }` + `color: var(--sp-red)` in
+   the inline style — but the reference deck didn't, so matching the
+   reference means literal colors). Recommendation: use literal colors
+   in the inline `style:` block (matches the reference deck's clean
+   approach) + update the test.
+
+**Confidence: 0.90** (pitfalls identified from the research findings
+above; mitigations are concrete).
+
+---
+
+## Assumptions logged (v1.23)
+
+| ID | Assumption | Confidence | Rationale |
+|----|------------|------------|-----------|
+| A-1.23.1 | The consolidated `-marp.md` will keep speaker notes as HTML comments (`<!-- -->`), which Marp strips from rendered HTML/PPTX but preserves in the source markdown for git readability. | 0.80 | Marp's markdown-it strips HTML comments by default; the source markdown keeps them. The alternative (drop notes entirely) loses the presenter context. |
+| A-1.23.2 | `python-pptx` will be added to `pyproject.toml [project.optional-dependencies] test` (not the runtime `dependencies`), since the structured PPTX is a build/test artifact, not a runtime dependency. | 0.90 | The render scripts run in CI + locally, not in the platform runtime. Matches the `moto`/`pytest` pattern. |
+| A-1.23.3 | The Marp PPTX (`.pptx`) is kept as a "fallback" alongside the new python-pptx structured PPTX (`-python.pptx`); both are committed + release-attached. | 0.85 | The plan says "dual PPTX"; the Marp PPTX is the quick render, the python PPTX is the editable one. Keeping both gives consumers a choice. |
+| A-1.23.4 | The `nova-sp-theme.css` file is KEPT in `docs/presentations/assets/` as a reference (retired from the render pipeline) but NOT deleted. | 0.90 | The plan says "keep as a reference (retired from render)"; deletion would lose the v1.22 work history. |
+| A-1.23.5 | The slide count after the word-count trim + "penetrate" removal will stay at 20 main + 1 appendix (the trim is within-slide prose, not slide removal). | 0.75 | The plan says "trim word count" (prose), not "remove slides"; but if a slide becomes too thin it may merge. To be confirmed in execution. |
+
+---
+
+## Decisions surfaced (v1.23 research)
+
+These research-grounded decisions will be bound in the CLARIFY stage:
+
+- **D-XXX (0.90):** Revert to `theme: default` + inline `style:` block
+  (port the reference deck's exact CSS from Finding 1). Keep
+  `nova-sp-theme.css` as a retired reference file. Confidence 0.90 —
+  the reference deck's clean look is the target; the CSS is read
+  verbatim.
+
+- **D-XXX (0.85):** Build `scripts/inline_images.py` (stdlib-only:
+  `base64` + `re` + `mimetypes`) to inline `<img src="assets/...">` →
+  `data:` URIs in the rendered HTML. No external library. Confidence
+  0.95 — confirmed the current HTML uses plain `<img src>`.
+
+- **D-XXX (0.85):** Build `scripts/render_pptx.py` using python-pptx
+  (>=0.6.23) for a structured, editable PPTX (`-python.pptx`). Keep
+  the Marp PPTX (`.pptx`) as a fallback. Both committed + release-
+  attached. Confidence 0.90 — python-pptx API confirmed.
+
+- **D-XXX (0.80):** Consolidate the plain `.md` + `-marp.md` into a
+  single `-marp.md`. Speaker notes move to HTML comments (stripped by
+  Marp from render, preserved in source). The plain `.md` is deleted.
+  Confidence 0.80 — the HTML-comment approach preserves notes for git
+  without leaking them to the audience.
+
+- **D-XXX (0.85):** No new personas for v1.23. lead-developer (deck
+  markdown + inline CSS + README + metadata) + backend-engineer
+  (Python scripts + bash + tests + CI). Territory enforcement = `warn`
+  for `render_pptx.py` (script is backend; slide layout decisions are
+  co-owned with lead-developer). Confidence 0.90 — follows v1.22 D-
+  148/D-149 precedent.