Fix selected outputs in the existing target eval cache

[jacobmichels]

A warm Tecnix target-cache hit can return the wrong output. For instance, the cache can return dev even when the target selected out. The existing cache remembers only the .drv path—the build recipe—so importing it loses the original output selection.

Reproduced on seven real World target IDs at commit 5ea0946b556c, using Tecnix 9d05607fe on x86_64-linux. Representative examples:

Target	Cold evaluation	Warm cache hit
//system/scm/worldtree:rustBin	out: executable binaries	debug: debug symbols
//docs/world:book	out: combined HTML/PDF directory	html: HTML output only
//system/aquifer/secrets-bootstrap:incidentLibgit2	lib: runtime library	dev: development files

Each cold/warm pair used the same World revision and .drv path, with an isolated cache and logs confirming the warm hit. The docs target’s Nix string dependency context also changed from out to html.

This PR fixes that existing behavior. It stores {drvPath, outputName} in a versioned result payload and reconstructs the selected output with its Nix string dependency information intact. Both fields are forced while source tracking is active, including a lazy output selector that does not change the recipe. Malformed payloads, unavailable recipes, and invalid output names fall back to evaluation.

The dependency-blob version advances from 1 to 2. Existing eval cache databases require no manual migration or deletion: old records become cache misses and are replaced as normal evaluation repopulates the cache. Users will see a cold cache that warms through use. This invalidation is necessary because older records never tracked output-selection inputs and cannot safely serve even dependency-only queries. The SQLite filename and schema, cache eligibility, evaluator-version handling, and packaging remain unchanged.

Checks:

• Full build and the repository formatter passed.
• Isolated Linux functional-test package: 215 passed, 0 failed, 10 skipped, with sandboxing enabled.
• Tecnix builtins and gc functional suites plus nix-expr-tests: 3/3 passed.
• Regressions cover selected outputs and string contexts, dependency-only warmup, deleted recipes, malformed payloads, lazy selector changes, and rejection of legacy dependency records.

Revision identities and build metadata for future remote-cache reuse are handled separately in #58. This PR depends on #59 for SQLite in the isolated test environment. #58 is the next layer above this selected-output fix. Native stack order: #59 → #57 → #58. This PR adds no remote transport or revision-stamping machinery.

[lilyinstarlight]

Perf testing shows no issues as well. I'm fine with this approach. I'll let you decide on the nit, I don't feel all that strongly on it

[lilyinstarlight]

Actually gonna leave one more comment, this does make one more cache miss/hit difference:

A cold target would have no outputSpecified value and a warm target would now have outputSpecified = true

The fix would probably be to make whether to add outputName to cache based on whether outputSpecified is present/true and fix why derivationToValue is returning a Value that has a default output that isn't the first output in outputs (at least according to PR description, I haven't personally validated this)

[jacobmichels]

A cold target would have no outputSpecified value and a warm target would now have outputSpecified = true

@lilyinstarlight I don't think this is true in our case. afaict nixpkgs does this, but I don't think our caching logic does.