fix: cost-allocation tag activation is backfillable; the stamp is not

[stxkxs]

The claim

Four places in cost-pipeline state that activating a cost-allocation tag is irreversible:

• README.md:96 — heading: "activation is not retroactive"
• README.md:111-113 — "Every hour before activation is permanently NULL for that key. There is no backfill."
• main.tf:776 — "Activation is not retroactive, so nothing before it is ever attributed."
• main.tf:1202 — "Spend inside that window is unattributable permanently"

And check-doc-contracts.sh required the README to contain the string not retroactive, so the claim wasn't just written down — it was held in place by a gate.

It's wrong

AWS ships cost allocation tag backfill: a management account can request up to twelve months, and it updates Cost Explorer, Data Exports and the CUR automatically on their next 24h refresh. One request at a time, one per 24 hours.

Reachable today in this account:

$ aws ce list-cost-allocation-tag-backfill-history
{ "BackfillRequests": [] }

What is actually irreversible

Backfill applies the tags' current activation status across history. It does not invent tag values. AWS states the constraint directly:

The resource tag must be historically assigned to the AWS Resource for the backfilled cost data to be available.

So the advice these docs give is right and the reason is not — and the difference decides what you hurry:

	recoverable?	cost of missing it
activating the key late	yes	one backfill request
resource ran untagged	no	that spend, permanently

Which is exactly why this component stamps PlatformId on its own buckets at install rather than waiting for the first tenant. That behaviour doesn't change; its justification becomes true.

The gate

A gate whose pass condition is a false statement would have rejected this correction. It now requires the README to say which half is recoverable.

Proven red by removing every occurrence of the word from the README and watching it fail, then green on restore. Worth recording that the first attempt at that proof was itself broken — the placeholder still contained the substring the check greps for case-insensitively, so the gate passed and looked vacuous when the test was the thing at fault.