fix(ci): pin pip <26.1 in Tiny AI workflow (lightning regression)

[lusoris]

Summary

• pip 26.1 (released today, 2026-04-30) regresses transitive resolution of lightning>=2.5,<3.0 — the pip install -e ai step in the Tiny AI workflow now fails with No matching distribution found for lightning<3.0,>=2.5 on every PR.
• Yesterday's PR refactor(gpu): land per-backend kernel scaffolding templates (CUDA + Vulkan, no migrations) #229 succeeded under pip 24.0 (lightning-2.6.1 resolved fine). PR feat(upstream): port d3647c73 — feature/speed extractors (speed_chroma + speed_temporal) #213 passed at 12:41 UTC today, then failed twice on reruns at 15:05 and 15:54 UTC after pip install --upgrade pip pulled 26.1.
• Narrow workaround: pin pip<26.1 only in the Tiny AI venv. Other workflows don't need lightning resolution and are left alone.
• Drop the pin once pip ships a fix release.

Six-deliverable checklist (ADR-0108)

• Research digest — no digest needed: trivial (1-line CI workaround for an external pip resolver regression; mechanism documented inline in the workflow comment + commit body).
• Decision matrix — no alternatives: only-one-way fix (the only narrow mitigation is to pin pip below the broken release; pinning lightning in ai/pyproject.toml does not work around the resolver bug).
• AGENTS.md invariant note — no rebase-sensitive invariants (CI workflow patch only; nothing rebase-relevant about a temporary version pin).
• Reproducer / smoke-test command — pasted below under "Reproducer".
• CHANGELOG.md — no changelog needed: CI-only change with no user-visible behavior delta. Per ADR-0100 §Per-surface minimum bars, infra-only patches don't ship CHANGELOG entries.
• Rebase note — no rebase impact: CI-only change against the fork's own workflow, doesn't intersect upstream Netflix/vmaf paths.

Reproducer

Failing path (current master, after pip 26.1 release):

python3 -m venv /tmp/venv
/tmp/venv/bin/pip install --upgrade pip
/tmp/venv/bin/pip install -e ai
# ERROR: Could not find a version that satisfies the requirement lightning<3.0,>=2.5
# ERROR: No matching distribution found for lightning<3.0,>=2.5

Fixed path (this PR):

python3 -m venv /tmp/venv
/tmp/venv/bin/pip install --upgrade 'pip<26.1'
/tmp/venv/bin/pip install -e ai
# resolves lightning-2.6.1 cleanly

Test plan

• Identified the failing job and step on PR feat(upstream): port d3647c73 — feature/speed extractors (speed_chroma + speed_temporal) #213 run 25172850010.
• Confirmed PR refactor(gpu): land per-backend kernel scaffolding templates (CUDA + Vulkan, no migrations) #229 (yesterday, pip 24.0) succeeded with the same ai/pyproject.toml.
• Watch this PR's own Tiny AI run — must pass.
• Once merged, rerun the Tiny AI job on PR feat(upstream): port d3647c73 — feature/speed extractors (speed_chroma + speed_temporal) #213 and confirm green.

Notes

No ai/pyproject.toml change — lightning>=2.5,<3.0 is a valid constraint that pip ≤25.x and pip ≥26.2 (once shipped) should resolve. This is a CI-infra mitigation, not a dep change.