chantier: Remotion gaps — the S-effort wave

[LeadcodeDev]

Follows the audit chantier (#142, merged by PR #163). The Remotion differential was re-scored against the post-chantier code: 1 gap closed, 11 started, 25 open, 6 re-scored, and 3 defects discovered while verifying.

Re-scored differential: https://claude.ai/code/artifact/5e71ffad-d959-441a-9448-c063c20ff7f2

Scope

The five S-effort items at the top of the ranking. They are independent of each other and split into 5 disjoint file sets.

Workstream	Subject	Value	Related issue
A	Distribute the 17 missing generation rules	High	#165
B	Finish relative-unit resolution (font-size + rollout)	High	—
C	Fix icon preloading, add a disk cache, make video decoding fail loudly	High	#166
D	Hardware-accelerated encoding	High	—
E	Spring with a pinned duration, and rest-time measurement	High	—

Out of scope

The two remaining Critical gaps — text measurement with auto-fit, and frame-range rendering — are M-effort and deserve their own chantier. The second one alone gates the whole distributed/serverless axis.

Issue #164 (freeze_at reapplied by hand in five render paths) is also out of scope here: its proper fix goes through the "generic time container" gap, scored M.

Verification contract

Same as the previous chantier, which proved its worth:

• Mandatory red phase with a numeric failure output for every fix.
• Each workstream verified in isolation on its own branch before merging into the chantier.
• cargo test --workspace, fmt --check and clippy -D warnings green.
• The 8 examples keep validating — 7 today, the eighth being examples/ferriskey-presentation.json does not pass rustmotion validate (pre-existing on main) #157, pre-existing on main.

For workstream A specifically, the expected fix is not to add the 17 missing files to the list, but to delete the manual list: that maintenance mode is what produced the defect, and it will reproduce it at the next rule added.

---

Outcome: delivered by PR #168. One deviation from the contract above: the workstreams were not verified in isolation — the orchestration saturated the machine and the agents stalled, so verification was redone sequentially on the combined tree. Details in the PR.

[LeadcodeDev]

Delivered by #168, merged into main. All five S-effort workstreams landed.

One deviation from the verification contract, recorded rather than glossed: the workstreams were not verified in isolation on their own branches. Giving each agent its own CARGO_TARGET_DIR multiplied a Skia-dependent build by five, the machine hit a load average of 75, and the agents stalled. Verification was redone sequentially on the combined tree, which surfaced three defects no agent had seen — including a call site left at 9 arguments after a signature grew to 10, meaning the tree did not compile. Details in #168.

Follow-on work: #169 (frame-range rendering) and #170 (text autofit) closed both remaining Critical gaps.