The Helm init container used /bin/RunProdServer which doesn't exist in IHP's
Nix Docker image (binary is at a Nix store path). Additionally, IHP v1.5's
RunProdServer starts the server after migrating — it never exits — so init
containers are the wrong pattern. IHP applies schema changes on startup.
Changes:
- Remove initContainers block from deployment.yaml entirely
- Set runMigrations: false as default in values.yaml
- Update RUNBOOK.md with correct skopeo push procedure (pre-fetch bearer token
to work around Gitea's misconfigured token realm URL: port 80 vs actual 32166)
- Add note that the Nix image has no /bin/sh or /bin/RunProdServer wrapper
k3s registry auth: credentials added to /etc/rancher/k3s/registries.yaml and
iptables DNAT rule added on Railiance01 (92.205.130.254:80 → 32166) so the
ACME token realm redirect works. hub.coulomb.social DNS A record still needed.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The inter-hub-lib postUnpack that expanded `import Generated.Types` to
119 individual imports was incorrect: it deleted `module Generated.Types`
from Prelude export lists without replacing it, so consumers of the
Prelude lost all entity types (Build 32: GHC-76037 not-in-scope errors).
Fix: keep Generated.Types as a real module in inter-hub-models (remove the
empty stub). With the ActualTypes.hi fix already in place (explicit T(..)
exports), the cascade is shallow: each entity .hi is compact, so
Generated.Types.hi stays well under GHC's 274 MB limit. This makes
`import Generated.Types` work normally throughout inter-hub-lib without
any source patching.
The entire inter-hub-lib overrideAttrs block is removed; the
inter-hub-models overlay now only rewrites ActualTypes.hs.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Root cause: libHSghc-9.10.3-5702.a (287,768,576 bytes) has its last AR
entry (Expr.o) claiming 517,544 bytes but only 82,258 bytes remain —
the archive is truncated. GHC's internal readAr (Data.Binary.Get) panics
at position 287,686,318 when it tries to read the full claimed size.
The truncated .a is read lazily: IHP's TH splices queue a dependency on
the ghc package, which flushes to readAr after all 477 modules compile.
This explains the invariant crash at [477 of 477] WidgetVersionInclude.
ghc-iserv-dyn is not exposed in ghc-with-packages/bin/ (why
-fexternal-interpreter alone silently fell back to the internal linker).
Use -pgmi with the absolute path in the unwrapped GHC store to force
iserv-dyn, which uses dlopen on libHSghc.so (intact, 110 MB) instead
of readAr on the truncated .a. No crash.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
-fexternal-interpreter had no effect: crash invariant at 287,686,318.
System ar reads libHSghc-9.10.3-5702.a (287,768,576 bytes) fine, but GHC's
internal readAr fails on the last entry — a bug in GHC's AR parser when
reading content near end-of-file. The call site is mergeObjectFiles during
.so creation, not TH evaluation.
--disable-shared skips the shared library build for inter-hub-models,
preventing GHC from ever calling readAr on libHSghc.a.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
GHC crashes at byte 287,686,318 reading libHSghc-9.10.3.a (~274 MB) via
its internal static linker during TH evaluation of WidgetVersionInclude.
-fexternal-interpreter delegates TH to a separate iserv process using the
dynamic linker, bypassing readAr and the 274 MB archive entirely.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
With -j8, GHC generates code for 8 modules in parallel. A parallel merging
step might read combined objects via Data.Binary.Get hitting 287 MB. Forcing
-j1 serializes codegen to test if parallel merging is the crash cause.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
DynamicToo state: DT_Dyn appears right before the crash. Force separate
static and dynamic GHC compilation passes to change the code path.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
